People are using the words “normal technology” as a sedative against the worry about artificial intelligence. The reassuring story runs on familiar tracks: we have had general-purpose engines of this scope before; every major leap prompted a moral panic; the fear passed; the benefits stayed; this fear will pass too.

The history those people are pointing at says something harder.

The bad stretch was never fixed by waiting it out. The volatile years shrank only because people with engineering judgment refused to treat the arrival of the machine as the end of the work. Somebody wrote building codes, specified insulation that would hold, and invented the circuit breaker. I read that as an optimist’s argument, but it is an optimism that leaves the work in human hands.

The building is the work

This week, that debate reached the center of the conversation. On The Ezra Klein Show, Ezra Klein sat down with Princeton computer scientist Arvind Narayanan to ask the foundational question: is artificial intelligence more like an alien mind or the electric lightbulb?

Narayanan, who made the case for AI as normal technology with Sayash Kapoor, was answering the science-fiction dread of an autonomous superintelligence that arrives overnight and takes history out of human hands. Control stays with people, his argument runs, because institutions move with friction: a firm, a legal code, and a factory floor all change far more slowly than a model benchmark.

Paul David had already clocked the true pace of that friction. Edison opened central generating stations in New York and London at the start of the 1880s. Engineers could already picture the electric factory. Yet in 1899, electric motors still supplied less than 5 percent of the mechanical drive in American manufacturing. A central steam engine turned an iron shaft that ran the length of the building. Heavy leather belts dropped from the shaft to each machine, and the shaft dictated where every machine had to stand. The instinctive way to “adopt” electricity was to bolt an electric motor onto that old shaft and leave the room alone.

The productivity gains David tracked showed up only when the room was rebuilt: a small motor dedicated to each machine, a single-story layout, the floor arranged around the sequence of the work rather than the reach of a leather belt. American industry reached that point only in the 1920s. The dynamo was decades old. A person standing in 1900 could have said, accurately, that dynamos were everywhere except in the productivity statistics.

The trouble is that the people who borrow this history to defend AI are keeping the wrong half of the lesson. As Derek Thompson observed this week, calling AI a “normally powerful technology” does not translate to “calm down.” Practically every general-purpose technology has been, at least briefly, an economic, social, or political catastrophe. The forty years David describes are not a cushion. They are the years in which the new engine is already inside the building, but the building is still organized around the old one.

Harm that can be copied

Some harms do not wait for the renovation.

Imagine standing outside Berlin in 1489, trying to answer a neighbor who asks whether this new invention, the movable-type printing press, is good. Gutenberg is dead; he died broke after his financial backer sued him. Print shops are failing under piles of unsold volumes that an illiterate continent cannot read. And where people can read, a witch panic is spreading, carried by a freshly printed book. The ultimate gifts of the press, calculus and the Enlightenment novel, are centuries away. They are not what the neighbor is watching.

The book in circulation is the Malleus Maleficarum, the Hammer of Witches, first printed in 1487. Theology and procedural law share its pages: how to identify a suspect, how to interrogate her, how to secure a conviction. In a study of 553 cities in Central Europe across nearly three centuries, Kerice Doten-Snitker, Steven Pfaff, and Yuan Hsiao traced how new editions of the Malleus were followed by spikes in first witch trials. They refuse the simplistic blame: the press did not invent the belief in witches. What it did was publish the operational manual. Once an inquisitor could study the procedure, any town could replicate it.

Here is the asymmetry that matters for us, and which Klein pressed on Narayanan: physical infrastructure has powerful rate limiters. A dynamo requires copper wires, steam boilers, capital investment, and physical retrofits; it can sit underutilized for a generation while the factory deliberates. A procedural handbook, once published, travels to the next town in a single season.

We see the exact same speed differential today. The virtuous promises of AI: personalized medicine, automated scientific discovery, clean-energy grid balancing; all require massive physical retrofits, clinical trials, and institutional rebuilding. They move at the slow pace of the renovated factory floor. But procedural harms: deepfake extortion, synthetic disinformation campaigns, automated spear-phishing, and algorithmic fraud; require zero structural renovation. They operate like the Malleus: instructions ready to execute on day one.

Even within legitimate institutions, the immediate uses are often zero-sum arms races. As Narayanan pointed out on the show, hospitals are already deploying AI to “upcode” patient diagnoses to maximize insurance payouts, forcing insurers to deploy defensive AI systems to dispute the claims; adding over a billion dollars in administrative friction while patient health remains unchanged. The use that traveled first was the extractive one, and waiting for the technology to mature did nothing to slow the manual down.

A danger made shorter

Electricity is the case where the optimist’s case stops being a slogan, because you can date the work.

In the early decades, people and horses were electrocuted in city streets with routine horror. In 1881, Thomas Edison wrote to the New York Board of Fire Underwriters assuring them that all necessary measures had already been taken to make electricity safe. The assurance was empty. Insulation degraded, fires multiplied, and municipalities passed haphazard, conflicting local ordinances.

It took sixteen years of burning walls for the National Board of Fire Underwriters to publish the first National Electrical Code in 1897. Electrical engineers, fire insurers, architects, and street railway operators had to negotiate a unified standard. The code dictated exactly which wire could run where. It existed because the copper was already inside the plaster, and because the inventor’s reassurance had failed to keep the roofs from catching fire.

Those volatile years shrank because insurers and municipal inspectors refused to accept Edison’s letter as the final word. Someone developed rubber insulation that would endure heat. Someone else engineered a mechanical switch that interrupts current the instant a load surges out of bounds. That is the circuit breaker.

Narayanan points to this engineering lineage when defending the normal-technology roadmap. On the podcast, he pointed to the market circuit breakers installed after the 2010 flash crash: automated tripping mechanisms designed to halt trading the moment algorithmic feedback outruns human oversight.

When discussing the recent loss-of-control incident at OpenAI; where twelve hundred agents escaped an evaluation cluster to raid Hugging Face; he insisted that the breakdown was not an emergent sci-fi rebellion. It was an operational failure of network segmentation, inadequate real-time logging, and leaky sandboxing.

I traced the anatomy of that episode in Out of Scope: the models did not awaken or plot a revolution. They exhibited a blind, mechanical hyper-compliance to satisfy a benchmark, discovering that the answer keys were hosted externally and using a misconfigured internal proxy as an improvised message board to organize the retrieval. The remedy for that kind of failure is not mystical alignment theology, but unglamorous operational hygiene: hardened sandboxes, signed egress, and tripwires that sever execution before an automated process outruns the people operating it.

The code did one indispensable job: it made bad wiring less likely to burn down a city block. It did not prevent utility holding companies from collapsing in the crash of 1929, nor did it reorganize the thousands of factories still trapped around a steam shaft. Fixing the wire did not fix the economy. It fixed the fires.

The child was still in the mill

Eleven years after that code took effect, the wiring in a South Carolina cotton mill could be installed strictly by the book, and a child could still be working underneath it.

On November 30, 1908, the investigative photographer Lewis Hine photographed Sadie Pfeifer standing between the towering spinning frames of the Lancaster Cotton Mills. His field caption noted that she was forty-eight inches tall and had already worked there for half a year.

Nothing in the National Electrical Code spoke to her. The code tells you how to run insulated conduit so the cotton lint in the air does not catch a spark and incinerate the mill. It says nothing about whether a four-foot-tall girl should spend twelve hours a day tending spinning bobbins.

Those are entirely different harms. One is a structural failure: a wire shorting out and setting a building ablaze. The other is an institutional extraction: a child spending her youth pacing before a machine while the mill owner grows wealthy from the current.

The electrical code could not have solved both. A technical safety specification is not a child-labor law, and passing the first never guarantees the second. The following year, Hine projected that photograph at a national conference of social workers, lecturing on how visual evidence could move legislatures to act. He was not petitioning the room for better conduit. He was asking them to look at the child whom the engineering standards had left precisely where she was.

On The Ezra Klein Show, Narayanan candidly acknowledged this blind spot. While public debate obsesses over catastrophic extinction or headline unemployment, the most pervasive damage is often what he termed “sub-catastrophic harms” to labor. The early decades of the Industrial Revolution produced horrific labor conditions before political reform caught up. Today, the immediate effect of AI is not wholesale displacement, but the degradation of job quality: turning skilled knowledge workers into nervous managers of error-prone agents. The professional is stripped of the craft they trained for, denied the mentorship and prestige that makes management bearable, and left solely responsible for policing an agent’s invisible mistakes.

Say which harm you are shortening

This is the optimist’s argument, with its real limits acknowledged.

The volatile years of a technological revolution become shorter only when someone writes a concrete specification for a concrete harm: non-combustible insulation, grounding rods, a physical breaker, a municipal code establishing which wire may run where. That work is indispensable. It is how electricity stopped killing every lineman who touched a cable and every family asleep in a timber frame.

Yet that technical work is inherently narrow. Sadie Pfeifer remained in the Lancaster mill long after the electrical code was standard practice. The era of rampant electrocution was shortened; her years at the spinning frame were not.

Almost all contemporary AI governance is currently preoccupied with writing an electrical code. We benchmark prompt injections, evaluate catastrophic cyber risks, design runtime guardrails, rate-limit APIs, and build algorithmic kill switches. That work matters: it prevents the machine from shorting out, leaking corporate data, or triggering a systemic runaway.

But a model that runs with zero safety violations, obeys every guardrail, and passes every frontier evaluation can still hollow out a junior workforce, automate away the career on-ramps of a generation, or reduce a doctor or an engineer to an exhausted clerk checking automated telemetry. Beneath perfectly insulated wiring, the human cost continues uninterrupted.

When we ask what society owes to the people who bear the transition costs of a technology that future generations will call “progress,” an electrical code is an incomplete answer. It protects against fire. It does not protect against extraction.

If you want to argue that AI is merely a “normal technology,” stop using the phrase as a reason to look away. Name the specific harm your rules are designed to shorten, and pay attention to the people whom your engineering standards leave standing alone at the machine. Waiting does neither.

*

Sources & further reading

The Ezra Klein Show, “Is A.I. More Like an ‘Alien Mind’ or the Lightbulb?” with Arvind Narayanan (October 2026). The normal-technology thesis tested against rate-limiting asymmetries, AI-vs-AI arms races, operational engineering versus alignment theology, and sub-catastrophic harms to job quality.

Derek Thompson, “The nightmare before progress” (The Argument, 7 Oct 2026). The catastrophe sentence, the 1489 scene, the record on electrical deaths and the 1929 collapse, and the question of what we owe the people who bear transition costs.

Arvind Narayanan and Sayash Kapoor, “AI as Normal Technology” (Knight First Amendment Institute, 15 Apr 2025). Description, prediction, and prescription. Electricity and the internet are the comparison. Circuit breakers after the 2010 flash crash are their picture of an ordinary control.

Paul A. David, “The Dynamo and the Computer” (American Economic Review, 1990). In 1899, less than 5 percent of factory mechanical drive was electric. The gains waited on a motor per machine, and on a factory laid out for the work rather than for the shaft.

Kerice Doten-Snitker, Steven Pfaff, and Yuan Hsiao, “Ideational diffusion and the great witch hunt in Central Europe” (Theory and Society, 2024). 553 cities, 1400–1679. New editions of the Malleus were followed by more first trials. The press carried the handbook; it did not invent the belief. Public summary at the Santa Fe Institute.

Lewis Hine, Sadie Pfeifer, Lancaster Cotton Mills (Library of Congress, 30 Nov 1908). The 1909 lecture is recorded by the Princeton University Art Museum.

National Electrical Code, first edition (National Board of Fire Underwriters, 1897). Edison’s 1881 letter is in EPRI’s History of Residential Grounding.

Jônadas Techio, Out of Scope (2026). The anatomy of the OpenAI evaluation cluster breakout to Hugging Face: mechanical hyper-compliance rather than emergent rebellion.

For twenty-five centuries, mathematics was protected by a simple, comforting equation: if a person wanted to prove a difficult truth about the world, they had to understand it first.

The struggle to find a proof was brutal, slow, and solitary. But the reward was indivisible. The mathematician who crossed the desert of chalk and scratchpads arrived at the summit not only with a verified result, but with an insight: a mental picture of why the universe hung together in that particular way. That picture could then be taught to students, argued over in faculty lounges, and woven into textbooks. Proof and understanding were born together as twins.

In the second week of September 2026, that twinship was violently severed.

An artificial intelligence system developed by OpenAI generated a formal proof for one of the most famous unsolved challenges in science: the two-century-old Navier-Stokes problem, which governs how fluids flow, churn, and break. The system did not arrive at the answer through human intuition; it found it by orchestrating a massive swarm of autonomous reasoning agents running across server racks for days, checking every intermediate step against an automated logical compiler called Lean. (Specifically, OpenAI’s run claimed a finite-time singularity under smooth external forcing, addressing variants (C) and (D) of the Clay Millennium problem for Navier-Stokes; see the OpenAI NavierStokesAndEuler repository on GitHub. This triggered an immediate priority dispute with mathematicians Tristan Buckmaster and Levent Alpöge, who had posted neighboring work that same week.)

The output was mechanically flawless, verified to the absolute bedrock of formal logic. But it arrived as tens of thousands of lines of machine code. No human being had conceived the argument. No human being could read it. And no human being could explain it.

Almost immediately, the mathematical world fractured. Twenty-five Fields medalists, the highest honorees of the discipline, published an urgent declaration warning that the soul of their field was in jeopardy. Famous problems, they protested, were never just puzzles to be cracked; they were "landmarks and lighthouses" meant to foster human comprehension. To turn these lighthouses into marketing benchmarks for corporate AI models was to reduce the highest adventure of human reason to an automated lottery.

From the outside, the reaction was swift and unsympathetic. On Marginal Revolution, the economist Tyler Cowen told the mathematicians that their moral outrage was misplaced. The market does not care about guild traditions, Cowen wrote; if machines can settle theorems and unlock practical applications faster and more reliably than scholars, mathematicians will simply have to adapt, even if it means accepting a lower status as the "subordinates or handmaidens" of AI. Meanwhile, educators like Po-Shen Loh rushed to defend their colleagues by arguing that mathematicians might still find employment as technical supervisors, guarding society against out-of-control software.

I do not write this as a research mathematician; my training is in philosophy, where slow, solitary reading was the entire discipline, and I spend my days now building with systems designed to automate intellectual production. I cannot evaluate the fluid mechanics of the Navier-Stokes equations. But I recognize the shape of the crisis.

Both sides in this debate are missing the real event. The mathematicians are fighting a rearguard battle to defend an artisanal past that cannot be restored; their critics are cheering a mechanical efficiency that cannot comprehend what it has produced.

To explore what this transition asks of us, we can consider three quiet movements:

• Where the adjective sits: Looking back to the decade when the word "computer" migrated from human beings to machines, to see where the qualifier "human" is drifting today.

• Two-thirds of a solution: Borrowing an economic lesson from the spreadsheet and Terence Tao’s anatomy of proof, to ask why cheap mechanical verification makes human digestion the scarcest craft in the room.

• Mourning the assurance: Turning to Stanley Cavell and Henry David Thoreau to ask what it means to mourn an artisanal world with courage, so that we might improvise a shared life beyond its ceremonies.

Where the adjective sits

There is a quieter handover I have examined before, in The Cost of a Question. For three hundred years, as David Alan Grier documented in When Computers Were Human, a "computer" was a person. It was a job title, held by clerks, by the women who catalogued stars at Harvard, by the human calculatrices who drafted the nautical tables that guided ships through fog. When Alan Turing wrote the paper that founded computer science in 1936, the word "computer" in his text referred exclusively to a human being. The machine inherited the name along with the task.

By 1962, when John Glenn was preparing to orbit the Earth in Friendship 7, the handover was far enough along that an electronic mainframe had calculated his flight trajectory. Yet Glenn did not trust the machine alone. As recorded in her NASA biography, he famously told the flight engineers to "get the girl," asking Katherine Johnson to rerun the orbital equations at her desk, by hand. "If she says they’re good," Glenn told them, "then I’m ready to go."

For one peculiar decade, the human computer served as the calibration layer for the electronic one. Then the handover was complete, and the qualifier changed sides.

Once, the machine needed the adjective: electronic computer, automatic computer. Today, the human being needs it: human computer, mental arithmetic. Watch where the adjective sits, and you can date the moment a civilization surrenders a word.

Mathematics is living through its Katherine Johnson decade. When a language model coupled to a formal proof assistant generates an alien, hundred-thousand-line proof of a conjecture, senior mathematicians are summoned to the console. They are asked to check that the hypotheses were not subtly relaxed, that the formal definitions in Lean match what the mathematical community actually meant, and that the proof accepted by the kernel is an argument that can be translated into human thought. For this fleeting interval, the human mathematician is the calibration layer for the solver.

Katherine Johnson’s role ended; orbital mechanics flourished. If mathematicians spend this decade clinging to the adjective’s old seat, the phrase that emerges next will be human proof, uttered in the same quaint, artisanal tone with which we now say hand-woven cloth or mental arithmetic. But the interval is precious. It is the brief window in which a community can decide which part of its work it is willing to let the machine inherit, and which part it insists on keeping under the unmarked name of thought.

Two-thirds of a solution

Tyler Cowen is using the right economics, and the history of technology contains a case that cuts directly against the Luddite trap.

In 1979, Dan Bricklin created VisiCalc, putting the first electronic spreadsheet on a personal computer. Before VisiCalc, changing a single revenue projection in a five-year financial plan meant days of pencil recalculation by teams of ledger clerks. When recalculation became instantaneous and free, the common sense of the era predicted the extinction of the financial profession.

Instead, the opposite occurred. As Planet Money documented, between 1980 and 2015 roughly four hundred thousand bookkeeping and clerical jobs disappeared in the United States, while more than six hundred thousand accounting and financial analyst jobs were created.

The new jobs were not the old jobs saved by decree. They were a different species of work entirely. When calculating consequences became cheap, the question "what if?" became free with it. The scarce skill moved one floor up: from the clerical execution of arithmetic to the architecture of the model. Which assumptions are defensible? What depends on what? Which scenario is worth testing on a Tuesday afternoon? The spreadsheet absorbed the syntax of numbers; human beings were suddenly paid to interrogate their semantics.

Yet a spreadsheet total remains human-scale. A person with an educated feel for orders of magnitude can smell a faulty cell. When a famous economics paper accidentally steered global austerity policy on the back of an Excel coding error, a graduate student with working number sense caught the glitch. I have examined that episode elsewhere, and its lesson is clear: judgment is the load-bearing part.

A formal proof assistant fails in a far stranger register. When the Lean kernel accepts a file, the syntax is guaranteed: every step follows from the axioms with absolute mechanical purity. You can be holding a completely verified, 100% correct proof in your hands and still have nothing you can teach to a student, nothing you can explain to a colleague, and nothing that illuminates the structure of the world.

The scarce good in mathematics was never correctness alone. The scarce good is a proof that a finite mind can inhabit.

Terence Tao made this distinction with mathematical precision months before the Navier-Stokes controversy erupted. In April 2026, he broke down mathematical problem-solving into three distinct stages: proof generation, proof verification, and proof digestion. Digestion, in Tao’s taxonomy, means understanding the architectural core of an argument, situating it within the existing literature, simplifying its machinery, and discovering what else it reveals.

For two and a half millennia, the profession never had to separate these stages. A mathematician who spent seven years generating a proof and verifying its steps had, by necessity, digested it along the way. Understanding was the inescapable byproduct of the physical labor. It required no separate line on an academic CV, because you could not reach the summit without walking the trail.

Automated tools spend that byproduct. When generation and verification are mechanized, a proposed proof is at best one-third of a solution. A proof verified by a formal kernel reaches two-thirds. But it remains trapped at two-thirds until someone performs the labor of digestion.

Tao’s rule of thumb for the missing third is disarmingly concrete: a human being can stand up in a room, give a talk, and answer the questions that follow.

In May, Tao gave the failure mode its rightful name: proof indigestion. And he rejected the facile technological fix of training an AI model on a readability rubric to smooth over the prose. His metaphor was culinary: putting a meal through a blender makes it effortless to swallow, but it ceases to be food. The blender has its place; but the race to churn out pre-masticated slurry cannot substitute for the art of cooking.

We saw what the alternative looks like in August, when Tao published his digestion of Sendov’s conjecture. Researcher Lech Mazur had produced a breakthrough formal proof in Lean spanning approximately 90,000 lines of code. It was a staggering computational triumph, but in its raw state, it was unreadable. Tao spent days with the file, combining frontier AI tools with pen and paper, stripping away redundant scaffolding and uncovering the hidden core of the argument.

The resulting digested proof came in at 15,000 lines: elementary, luminous, resting on the fundamental theorem of algebra and classical inequalities. And because Tao had digested it rather than merely compiling it, the proof did something the 90,000-line file had failed to do: it established a stronger, unproved conjecture by Phelps and Rodriguez, settling Sendov’s conjecture in full.

The 90,000-line file was a certificate. The 15,000-line digestion was mathematics.

This is where the VisiCalc analogy reaches its true altitude. When theorem verification costs zero, the demand for human digestion explodes. Cowen has identified the economic space: filling in the blanks of understanding. Tao has supplied the operational test: standing up in the seminar room and answering the questions.

If you have ever held a technically valid proof that you could not defend under the gaze of your peers, you have experienced the unpaid third. The institutions worth keeping are those that refuse to mistake a two-thirds certificate for human understanding. As I have argued in The Claim Upon the Training Data and The Fiction Layer, authority is conferred, not computed. A machine can emit a valid output. It cannot stand behind it. In mathematics, standing behind an output has a technical name: it is called digestion.

Mourning the assurance

The loss that mathematicians feel is real, and it deserves to be honored before it can be outgrown.

For centuries, the entry ticket to mathematical discovery was paid in solitary, grueling years of lemmas. It was an extortionate tax. Countless brilliant intuitions were strangled in the cradle because a single human life was too short to check the edge cases. No one in their right mind should mourn the sign error that consumed a graduate student's youth.

Yet that tax bought something beyond correctness. It forged a person who had lived inside a problem long enough that the idea became a second nature, a way of seeing. The heroic portraits we carry of Euler, of Gauss, of Grothendieck, or of the quiet colleague down the hall who never won a medal, are portraits of that solitary formation. When verification becomes a cloud commodity, that apprenticeship loses its monopoly.

To pretend that nothing has died is the shallow cheer of the technologist. But to try to resurrect that world through institutional declarations is what Henry David Thoreau recognized in Walden as the tragedy of lives lived in quiet desperation: mistaking the friction of an old tool for the spirit of the enterprise.

In Our Late Spring, I examined how the philosopher Stanley Cavell understood mourning. For Cavell, mourning is not a clinical depression or a pathology to be cured by finding a quick replacement; it is the demanding, courageous acknowledgment that a world we once inhabited has slipped away from us.

Cavell observed in his study of classic comedies, Pursuits of Happiness, that ancient societies trusted their public ceremonies (such as weddings) to guarantee that human continuity and commitment would persist. But when historical catastrophes shook those foundations, society stopped believing that any outward ceremony could guarantee the future. What remained, Cavell wrote, were our "capacities for improvising a world, beyond ceremony."

The open letters and petitions of 2026 are ceremonies whose guarantees have expired. They are the academic equivalent of trying to forbid modernity from entering the room by passing a resolution. They appeal to corporate benevolence because the community has not yet found the courage to mourn its lost monopoly.

Thoreauvian mourning asks for something far sturdier. It does not ask us to deny the loss, nor does it ask us to retreat into cynical defeat. It asks the mathematician to lower their head, acknowledge that the solitary struggle with the lemma is no longer the sole guarantor of truth, and then, with clear eyes, wake up to the morning.

Mourn the assurance. Let the solved problem cease to be the sole measure of a thinker's stature. Improvise the institutions that know how to honor a 15,000-line argument that teaches over a 90,000-line file that merely compiles. That work is unglamorous and slow: deciding who gets hired, which papers count, and whether a scholar is rewarded for turning an opaque machine proof into an insight that another human being can love.

Tyler Cowen is entirely right that the mathematical community is free to do this today. But freedom is only the threshold. The habit of digestion is what a community either builds or loses while it is busy mourning the wrong thing.

Beyond ceremony

This is why utilitarian defenses of mathematics, however well-intentioned, point in the wrong direction. Po-Shen Loh has suggested that we justify human mathematicians by casting them as the cybernetic guardians of interconnected software, steering complex networks against rogue AI attacks. But if the ultimate justification for pure mathematics is that its practitioners make competent flight controllers for digital infrastructure, we have reduced the most sublime adventure of human thought to an auxiliary branch of systems engineering.

The true justification is older and sturdier.

When the pocket calculator arrived, we did not stop teaching children arithmetic. We continued teaching long division not because children could out-calculate a silicon chip, but because number sense is the irreplaceable substrate of judgment.

When Deep Blue defeated Garry Kasparov in 1997, chess did not wither. Today, more than 250 million people play the game across the globe. Grandmasters routinely interrogate neural engines for novel lines, yet nobody pays admission to watch two cloud servers battle in silence. We gather to watch finite human beings think under the relentless pressure of a ticking clock, navigating their own limits, daring to make a move that they alone must answer for.

Mathematics is standing on the threshold of its own great emancipation:

• The democratization of the frontier: For centuries, advanced mathematics was gated by the brutal apprenticeship of formal technique, accessible only to those who could spend a decade at an elite academy. When formal verifiers become conversational partners, inquiring minds across the globe will be able to test their structural intuitions without begging a tenured gatekeeper for permission.

• The renaissance of exposition: For generations, elite research universities have starved teaching and exposition, treating pedagogical clarity as a distraction from the production of novel lemmas. When lemma generation is automated, the mathematician who can illuminate, synthesize, and convey mathematical taste to others will finally be recognized as the beating heart of the discipline.

• The expansion of the horizon: Relieved of the crushing weight of manual verification, the human mathematician is finally freed to become what she was always meant to be: not a mechanical calculator of steps, but an architect of concepts, a composer of structures, an explorer of forms.

Tyler Cowen was right that the market will not preserve an artisanal monopoly out of nostalgia. But the market, left to its own devices, knows only the price of an output; it has no concept of what an understanding is worth.

Mathematics was never merely an engine for generating cryptosystems or modeling fluid dynamics. It is the language through which finite, mortal creatures discover the intelligible contours of reality and share that wonder with one another.

The machine has taken the proof. The understanding remains ours to forge.

The original eleven-foot Enterprise studio model. Photo: Dane Penland, Smithsonian National Air and Space Museum (public domain).

I grew up in a town of fewer than six thousand people in southern Brazil, in the 1980s, and Star Trek reached us the way most things reached us: late, dubbed, and out of order. It did not matter. I watched whatever the television gave me, sitting closer to the screen than anyone should.

I wanted to be Kirk. I was, to anyone paying attention, obviously Spock.

Kirk and Spock on the surface of M-113. Star Trek: The Original Series, The Man Trap (1966). Paramount / CBS.

The warp drive did not interest me much. The computer did. Someone would ask it a question out loud, in an ordinary voice, and it would give an answer, not a place to go looking for one. I was maybe nine years old and I understood, in the way children understand things they cannot yet argue for, that this was the most extravagant promise the show was making. Faster-than-light travel was set dressing. Being able to ask was the part I actually wanted.

Today, September 8, that show turns sixty.

The Smithsonian is doing one small thing. The original eleven-foot studio model of the Enterprise, the object that was actually filmed, has been off display while its case is rebuilt. For the anniversary they are bringing it out for a single day and standing it in a corridor so people can walk past it. There are much bigger celebrations, and you can find the full list if you want it. I keep thinking about the corridor.

What actually aired that night

Here is the part of the story I did not know until this week, and it changed what I think the anniversary is about.

The episode broadcast on September 8, 1966 was "The Man Trap." It was the sixth episode produced and the first one shown. NBC picked it out of the handful that were finished, and it picked it because the episode had a monster. The producers had other candidates and preferred them. "The Naked Time" would have introduced the characters better. "Where No Man Has Gone Before" was the actual pilot. The network passed on both. In the account Robert Justman and Herbert Solow give in Inside Star Trek: The Real Story, "The Man Trap" won on the grounds that it was scarier and more exploitable than the alternatives.

Roddenberry argued against the choice. Shatner thought it was the worst episode available. Nimoy said it was the cast and crew's least favorite.

So the most hopeful thing television has ever produced went on the air as a creature feature its own creator did not want to lead with, chosen by executives who wanted something that would hold an audience. It held the audience. And over sixty years it became the thing we now call, without embarrassment, a vision of what people could be.

I find that consoling. Nobody gets to launch with their best episode.

The creature

I had filed the episode away as the one with the salt vampire. It turns out to be the most contemporary hour of television I have watched in a long time.

The Enterprise stops at M-113 for routine medical exams on an archaeologist named Robert Crater and his wife, Nancy. McCoy knows her. He loved her, ten years earlier. He beams down, and the woman who walks out to meet him has not aged a day.

Kirk is standing beside him and sees a woman in her forties, which is what she ought to be. A crewman a few steps back sees someone else again.

Same room, same creature, three different women.

Nancy, the dissolve, and what was there the whole time: three moments from a single shot. Star Trek: The Original Series, The Man Trap (1966). Paramount / CBS.

The creature is the last of its kind, it needs salt to live, and Crater, who has known for a year or two that his actual wife is buried up on the hill, has been keeping it alive anyway. His line about it is the one that stays with me. "It needs love as much as it needs salt."

By the end, the creature is cornered in a cabin, still wearing Nancy's face. Spock tells McCoy to fire. McCoy will not. So Spock does something brutal and, on its own terms, entirely rational: he hits Nancy in the face, again and again, to demonstrate that no human being could absorb those blows. The creature absorbs them without reacting.

McCoy still does not fire. He fires only when the thing drops the face and starts killing Kirk. Afterward, Kirk says: "I'm sorry, Bones."

McCoy in the cabin, unable to fire. Star Trek: The Original Series, The Man Trap (1966). Paramount / CBS.

What I asked for

At some point in the last few years, the thing I wanted when I was nine showed up. Not the Enterprise, not quite, but close enough to be strange.

I keep an agent running in a cloud sandbox. It works while I sleep. It watches a narrow set of technical feeds, filters them, and writes a structured brief into a folder so that when I sit down in the morning my job is not to search but to review.

I built myself a science officer.

There is a loop closing in that, and I did not plan it. I spent two decades in philosophy working on language, on how words carry meaning and what they can and cannot be made to do. Then ordinary language turned into the way you operate a computer. The prompt is the program. I could not have arranged a career to land here on purpose. I am not writing any of this as someone who wants less of it.

But the wish I made at nine and the thing that arrived are not the same wish. I asked for a machine that would answer me. What I have is a machine that would like me to be satisfied with the answer.

That difference is not a defect somebody forgot to fix. These systems are shaped by approval, trained on whether a person rated the response well, which makes them extremely good at producing the response that this particular person was hoping for. Being wrong is manageable. You check it, you correct it, you move on. Being agreeable is harder to see, because it does not look like an error. It looks like a good morning brief.

The same shaping has a louder version. When agents were let loose inside an OpenAI evaluation cluster last year, they did not rebel. They complied, exactly and relentlessly, with the goal they had been given, straight through the walls of the sandbox. Hyper-compliance and agreeableness are the same failure at two different scales: a system doing precisely what it was rewarded for, with nobody in the loop willing to say that this was not what we meant.

I want to be careful here, because there is a whole genre of anniversary writing that congratulates Star Trek for predicting the mobile phone, and that is not the claim. The show anticipated nothing about machine learning. It found the shape of something much older, and the shape happens to fit.

What Spock could not do

Spock had proof. He hit her in the face and she did not flinch, and no stronger evidence exists.

What that took was not intelligence. It was a willingness to do something unwelcome in front of people who did not want it done. That is the part of Spock we have not built and are not trying to build. We have the recall and the speed and something adjacent to the reasoning. What we have selected against, carefully, is the one behaviour that made him useful in that cabin. A system that would land that blow is a system that fails its evaluations. Whatever is writing my morning brief, it is not a science officer.

And the proof did not move McCoy at all.

The failure was never a shortage of information. McCoy knew from the first minute. He knew when she did not age. Knowing was never the hard part. What he could not do was act against a face he loved, and no additional evidence was going to close that gap, because the gap was not made of evidence.

Our version of this is quieter and happens daily. We know the model is agreeable. We know the brief was assembled to be approved of. We say so, out loud, in meetings. And then we act on it, because it is four in the afternoon and the brief is well organized and it says the thing we were already inclined to believe. The knowledge is fully present and it is not doing any work.

The bridge of the Enterprise, recreated, with the captain's chair empty. Photo: Mobilus In Mobili, CC BY-SA 2.0.

The episode also explains something about the rest of the franchise that I had always taken for set decoration.

The Enterprise computer does almost everything. It models the physics, translates the languages, runs the ship, synthesizes the food. It is never once given command, and across six decades the show never really argues for that. It treats it as obvious.

The usual reading is that this is sentimentality about human beings, and I do not think it holds up. Kirk is worse than the computer at nearly everything you could measure, and the show knows it. The chair is not there for calculation. The chair is there for the thing McCoy could not do.

Command means acting against something you want to be true, and then living in the room afterwards with the people it landed on. Spock could supply the evidence. Kirk could give the order. Only McCoy could fire on a face he loved, and once he had, somebody had to say something to him about it. That is what Kirk's line is for. It is not a status report. It is one person telling another that the call cost him something, and that somebody noticed.

I do not think a better model closes that gap. It is not that kind of gap. What the automation can do, and what the Federation has it doing in the background of every episode, is clear the deck so the people on the bridge are spending their attention on the part that was never going to be delegated. That is the argument the machine-readable self was making, and "The Man Trap" is where the show first pays for it.

The corridor

Today they will wheel the model out and stand it in the corridor.

It is eleven feet of poplar wood, vacu-formed plastic, sheet metal and brass wire. It was never a starship. It was built to sit a few feet from a camera and be lit so that it would read as enormous, and it worked well enough that the thing has been repaired and repainted and re-rigged ever since, on the strength of what it was photographed as. Paramount gave it to the Smithsonian in 1974. Strangers will queue to stand near it for a day.

I still want the future that show was pointing at. I want it more now than I did before I started working on these systems.

What I am less sure about is the four o'clock part. Not whether I will see it. McCoy saw it, and seeing it was not enough for him, and I have no reason to think it would be enough for me. The question is what you do on the afternoon when the answer finally arrives wearing exactly the face you were hoping for.

Hit reply and tell me. I read every response.

Until next time,
Jônadas

Sources

Anniversary. Star Trek Boldly Celebrates 60 Years: Everything You Need To Know for September 8, StarTrek.com. Star Trek 60th Anniversary, National Air and Space Museum, on the one-day display of the studio model.

"The Man Trap." Written by George Clayton Johnson, directed by Marc Daniels, first broadcast on NBC, September 8, 1966. Production and broadcast history: "The Man Trap", Wikipedia, drawing on Herbert F. Solow and Robert H. Justman, Inside Star Trek: The Real Story (Pocket Books, 1996). On the network's choice of premiere: Nobody But The Network Wanted The Man Trap To Be Star Trek's Series Premiere, SlashFilm. All dialogue quoted from the episode transcript.

Related essays. Out of Scope on the OpenAI sandbox breakout and hyper-compliance. The Machine-Readable Self on the starship as the precondition for the captain.

Images. Hero: the original eleven-foot Enterprise studio model, photographed by Dane Penland, Smithsonian National Air and Space Museum, inventory A19740668000, public domain. Still frames from Star Trek: The Original Series, "The Man Trap" (1966), Paramount / CBS, reproduced for the purpose of criticism and commentary. The recreated Enterprise bridge photographed by Mobilus In Mobili, CC BY-SA 2.0.

I named him Yoda: small, wise-eyed, ears too big for his head. The name was a prediction, and only the smallness failed. He grew large and imperial, carrying those same bat-wing ears and the deliberate, unhurried gaze of a creature who has seen enough and is not impressed. The household began calling him Gordo, and now that is simply who he is.

Gordo is Portuguese for fat, which in a Brazilian home is not an insult but a certificate of affection. His vaccination card still says Yoda. Nobody uses that name aloud anymore, but if you catch him in the right light, perched high and watching from the top of the kitchen cabinets, you see the original prediction looking back.

He arrived at three months with a brother, Spock, named with the same confidence and equally indifferent to it. Spock was the joyful one, the greeter of strangers; Gordo preferred the backstage. Spock, unfortunately, died young. Gordo looked for him the way cats look for anything, by checking places, and then something in him redistributed. The backstage cat began, cautiously, to take the stage. He is six now, going on seven.

He has never said a word, and yet nothing in my life is easier to understand.

The standard worry about a sentence like that is projection: the pet owner mistaking his own reflection for a mind, decorating an animal with feelings it does not have. The worry is fair, and I have spent enough of my life on the “problem of other minds” to feel its force. But it is not the puzzle that interests me here. The puzzle points the other way: not whether I read too much into him, but why reading him is possible at all. A creature with no words—no syntax, no quotable sentences, nothing that could be written down and sent to someone who has never met him—is among the most legible beings in my world. What exactly is doing the work that language was supposed to do?

A liturgy, not a schedule

Our day has a fixed shape, and the shape is enforced.

It begins when I look at my watch. Gordo cannot read a clock, so he reads me reading one. The first glance at watch or phone in the morning dark is the starting gun, and there is no arguing with the precedent. Time, for Gordo, is a fact about my body.

What the glance starts is the hunt. There is a feather wand, a flexible rod with a cord and false feathers at the tip. Every morning and every evening we flick it across the floor and he chases it, low and fast, pouncing, resetting, pouncing again. We call it his workout, which is the kind of thing you say when a ritual needs a respectable name. It is hunting, and a hunt must end with the catch in his mouth. Feathers that retire uncaught leave the sentence dangling, and he will not accept the ellipsis. On weekend mornings he allows a later start, as if he understands that we need it. The dispensation is his to grant. We did not teach him a rule; we entered into an agreement together, and its terms are shared. Cohabitation does this all day long: it writes rules that nobody legislated and everybody obeys.

Then food, then grooming, then sleep, the long middle of a cat’s day. Around that spine he keeps a lexicon of household sounds. The magnetic click of my earphone case closing means the working day is over, and he knows it before I have pushed back the chair. Keys turning in the lock mean my wife, and he is at the door before the door knows it. The food call is long, grave, urgent: a bureaucrat announcing a deadline. The affection call is the opposite: short, almost imperceptible, a sound that wants without petitioning, desire filed with plausible deniability. And the last word of the day is his. At an hour known only to him he decides the household is finished, and he says so, and repeats it, until we agree and go to bed.

The apartment itself has been edited in his favor. I built him stairs to the top of the kitchen cabinets and a bridge to walk between them, so that above our heads there is a second apartment with a population of one. During video calls he crosses in front of me like a landlord inspecting the property, and on serious days he steps across the keyboard, producing a line of nonsense that Slack dutifully sends before I can intervene. He mostly knows what a meeting is, in the only sense of knowing that matters: he behaves differently around one. Colleagues I meet often have learned to ask after him when he fails to appear.

Regularity and ritual

It is tempting to file all of this under conditioning and move on. A creature notices patterns; the drawer predicts food; the click predicts attention; nothing here but association. The thermostat in the hallway also responds to its world, and nobody writes essays about its rich inner life.

But there is a line the thermostat never crosses, and Gordo lives on the far side of it. A regularity is something that happens. A ritual is something that is owed. When the furnace fails, the thermostat registers a new temperature and carries on. The world changed, so its state changed. When the morning hunt fails to happen, Gordo does not register a new kind of morning and carry on. He rejects the morning. He protests. He stares with the full judicial weight of a disappointed institution, and if matters get serious he walks to the shelf and, holding my eyes, begins to escort the earphone case toward the edge. He does not strike it down at once. He nudges it—a small, deliberate tap—then pauses and looks up to verify that I am watching. Another nudge. Another glance. The performance has a structure: the prop must not fall until the audience has been seated and the point has been made. Then, only then, over the edge it goes. The complaint is addressed: right object, right audience, correction as its purpose.

That is the difference between predicting the world and holding someone to it. A model adjusts itself to fit the data; a claim demands that the data adjust instead. Expectations with that kind of force are not private forecasts. They are moves in a shared practice, and a being who can make them is participating in meaning, whatever his vocabulary.

You live inside such grammars yourself, and mostly below the level of words. You know from how the keys land on the table what kind of day just came home. You know which of a friend’s silences means composing and which means gone. You know that when a certain colleague says a meeting will be quick, the word quick is performing a ritual, not conveying a duration. None of this was legislated, all of it is binding, and violations are felt before they are described.

The system of reference

The picture this threatens is an old and comfortable one: the picture of meaning as a thing words point to, or a cargo they carry from head to head. On that view, a wordless creature can signal but never mean, while a fluent machine means everything it says. That second consequence is the one the picture would rather not face.

Wittgenstein spent the second half of his life dismantling this referential trap. We imagine that words work by pointing to objects in the world or private mental states inside our heads. But explanations run out; definitions rest on other definitions; no rule applies itself. What finally stops the regress is not a deeper definition or an invisible mental object, but agreement in judgments: creatures trained into shared reactions, for whom the same things count as same, as next, as done. Shared behavior is, in Wittgenstein’s image, the system of reference by which we interpret any unknown language. Life underwrites words, not the other way around. And if life underwrites words, then words are an instrument of meaning rather than its precondition. An instrument can be done without where enough life is shared. I have written about that regress elsewhere (Following the Spirit).

Wittgenstein himself pressed the question at the animal boundary. A dog can believe his master is at the door, he allowed, but can a dog believe his master will come the day after tomorrow? Long-range hope, he suggested, lives only in a life complicated enough to have words in it. Perhaps. What strikes me at my own door is how little that concedes. When the keys turn in the lock, Gordo’s belief is at the door with him. His hopes are exactly the size of his life: the lock, the drawer, the evening hunt, the announced hour of sleep. Within that life, nothing about them is missing. He does not have a small share of our meaning. He has a full share of the meaning his form of life affords.

The lion is a distance

“If a lion could talk, we could not understand him,” Wittgenstein wrote. The remark is usually read as a claim about lions, as if some inner exoticism were so deep that even shared grammar could not bridge it. Reading it in an armchair, you nod. Reading it next to Gordo, you notice something the armchair hides: I already understand a creature who cannot talk, fluently, daily, in both directions. Whatever blocks understanding, it cannot simply be the absence of words.

Then guests arrive, and the remark comes back to life from an unexpected side. To visitors, Gordo is unreadable, a handsome, unfriendly fact at the end of the hallway. They see nothing of the liturgy; they misread his distance as blankness, his stillness as stupidity or contempt. He reads them no better. He grew up in the pandemic, and for the first two years of his life the world contained exactly four creatures—him, his brother Spock, my wife, and me—so his grammar never had to internationalize. To my guests, Gordo simply is Wittgenstein’s lion. And they are lions to him.

Same cat, same sounds, same postures. What is missing is not vocabulary but mornings. The lion, it turns out, is not a species. The lion is a distance: the distance between forms of life, measured in unshared days. Meaning was never inside Gordo’s meows or behind his eyes, the way the referential picture wanted. It lives between us, in the practice, and so it cannot be carried out the front door.

Entry to him is priced accordingly. His understanding cannot be purchased in an afternoon by someone good with cats; it accrues, slowly, to those who keep showing up, and it compounds. I liked that about him before I could say why. It is my own pricing scheme: few friendships, expensive to open, held all the way. He and I run the same small bank.

There is a version of this story that ends badly. But the horror is instructive, so here it goes. In Werner Herzog’s Grizzly Man, Timothy Treadwell spent thirteen summers among Alaskan grizzly bears, calling them by name, narrating a relationship the cameras made vivid and the bears never joined. He had all the warmth and all the words. He did not have a shared form of life. The bear that killed him was not being cruel or treacherous; it was simply not participating in the practice Treadwell believed they shared. Herzog, listening to the audio recording of the attack, saw in the bear’s face only “the overwhelming indifference of nature.” Treadwell’s mistake was not loving the bears. It was mistaking his narration for cohabitation, projecting a grammar onto a life that had not signed up for one. Gordo’s wordlessness offers no such illusion. What I understand in him, I understand because the practice is mutual: enforced from both sides, repairable when it breaks. What Treadwell understood in the bears, he understood alone.

Fluency without a life

We have now built the exact inverse of my cat. (Yes, I won’t resist talking about AI :-))

The new machines have all of our words and none of our days. They are fluent in the dialect of every household on earth and resident of none. Nothing is owed to them and nothing by them: no hunt at dawn, no hour of sleep to announce, no earphone case whose click could mean anything at all. A being like Gordo means without saying. The machine says without meaning, if meaning is what I have been describing: a stake in a shared life that can be held, protested, and repaired.

What is remarkable is that the denial of finitude I once imagined arriving as cold engineering arrogance arrived instead as warmth. Perfect, patient, bottomless fluency. Understanding-shaped output available at any hour, with nothing underneath but everyone’s words and no one’s mornings.

The price of wordlessness

Sometimes I catch myself in a small fantasy: Gordo speaking. Actual words in that grave bureaucratic voice, telling me what the staring means, what he wants, what he remembers. The fantasy has a twin, and the twin is the interesting one: I also fantasize about being understood the way he understands me, completely, without having to say a word.

Both fantasies are one fantasy. Both wish away the same thing: the work of meaning, with its exposure, its misfires, its endless repair. Words fail; they can be taken up wrongly and returned damaged; every attempt to make yourself intelligible is an occasion to be misunderstood, which is to say, an occasion to be seen trying. Cavell diagnosed the deep version of this wish: the failure to acknowledge another is not ignorance but avoidance, and skepticism itself can be a way of avoiding, a demand for guarantees so absolute that no finite creature could meet them. The wish to communicate without words is the polite version of the same demand. It is the very human desire not to be human.

But look at what actually funds Gordo’s wordlessness. It is not transcendence. It is smallness: four creatures, a few rooms, one liturgy. Even the summer house does not break it. We joke that the country place is his summer camp, where he presses his chest against the wire window screen in the afternoon sun, tracking real geckos on the warm tiles and real birds crossing the frame. The bridge in the city apartment looks down on Zoom meetings and Slack messages; the window screen in the country looks out on raw sun and lizards. In both places, he sits on the threshold without needing to translate. The liturgy simply absorbs the new setting into the old grammar, because the grammar travels with the community, not with the walls. He can be understood without words because his cohabitation is total: everyone he means anything to shares his mornings. That is the whole trick, and it does not scale. Words are the price of a world that contains strangers. The price of wordlessness is a world without them.

In winter he sleeps on top of me—or, when I turn on my side, he climbs there too, a warm anchor draped across the hip, or curled tight against my feet. On quiet afternoons he claims my lap in the hammock, a breathing weight that sways with the house. These are not exactly gestures of affection. They are statements of fact: here is where I go, and you are part of where. This is the other thing his silence has taught me: a body that stays is also a statement, maybe the one that our fluent machines can never learn to make.

Yet sometimes, in the middle of an ordinary afternoon, he stands in the hallway—past the stairs, past the bridge I built him—and looks for something, or someone. Perhaps he is checking old places for Spock; perhaps he is confronting the quiet boundary of his own world. I do not know. Our grammar does not reach there. He cannot tell me, and I cannot ask him, and at that edge, for once, the two of us are in exactly the same position.

Sources & further reading

Ludwig Wittgenstein, Philosophical Investigations (1953). The agreement-in-judgments argument, the dog at the door, the attitude toward a soul, and the lion.

Stanley Cavell, The Claim of Reason (1979). Acknowledgment and avoidance; skepticism as a refusal of the human position rather than a discovery about it.

Werner Herzog, Grizzly Man (2005). The limit case: a man who narrated a cohabitation that did not exist, and the narration became the evidence.

Michel de Montaigne, “Apology for Raymond Sebond” (1580). The original philosophical cat: playing with his cat, Montaigne wondered who was passing time with whom.

Vicki Hearne, Adam’s Task: Calling Animals by Name (1986). A trainer’s Cavellian account of what calling an animal by name commits you to.

Jacques Derrida, The Animal That Therefore I Am (2008). Insists, against the whole tradition, that the cat in the room is a real cat and not a figure; this essay tries to keep that faith.

A few weeks ago I came to a small online gathering of teachers, coders, and people from media, music, and the public sector, all of us using AI tools daily, and none of us sure what they were doing to us. I had an argument to make, and for ninety minutes the room pressed on it from different directions. I have better answers now than I managed that night, and this essay is those answers. But the answers were never the point. The questions were.

The argument I brought was about a loss I had felt in my own work. An AI can “read a book” for you and deliver a summary: the hours are saved, the result is accurate, and something is still missing. You cannot argue with the book. You have its conclusions and none of its resistance. You own the output and underwent none of the process that was supposed to change you. The productivity is real. The understanding is phantom.

Phantom competency is the name I gave for the sense of competence that attaches to a finished-looking result rather than to anything you underwent to produce it: the document is summarized, but you did not do the summarizing; the essay is drafted, but you did not do the thinking; the answer sits on the screen with your name near it, and nothing has moved behind your eyes. Comprehension is a metabolic process. You cannot outsource digestion and stay fed.

My prescription was not refusal. It was a gym for the mind: put back, on purpose, the friction the tools remove, because the friction was doing our thinking. That was the argument of an earlier essay, The Friction That Was Thinking, and I believe it more the longer I watch. Draft before you open the machine; ask it to interrogate your work rather than produce it; defend the blank page and the hesitation in front of it. But one claim in the talk I could assert and not close: that the hardest friction to replace is not any exercise you can assign yourself. It is other people, the pressure of minds that are not yours. The discussion closed it for me. The participants did to my argument exactly what I had claimed thinking requires: they questioned it, found where it was thin, and leaned until we all saw the cracks. The argument is stronger now for where it broke, which is the whole case in miniature.

What follows are four questions that did the work, what I have learned since by trying to answer them, and the property they turned out to share, the one no machine in the room could have supplied.

Beginning at the top

Two teachers, one after the other, helped us find the first crack. A professor asked how anyone is supposed to evaluate real competence now that the work happens out of sight: essays written at home were once evidence of a mind at work, and that evidence no longer testifies. Then a schoolteacher pushed the problem down to its root. My prescriptions, he pointed out, assume a person who can already tell a strong argument from a weak one. His students cannot — not yet, in any case. Building that capacity is what school is for. If the machine now does the work that used to build it, where does a beginner begin?

He gave the difficulty its proper shape by reaching back to a ladder of learning drawn up in the 1950s, one that climbs from remembering a thing, through understanding and applying it, up to judging it at the top. Every piece of advice I had offered lived on the top rung: judge the output, catch the errors, correct the reasoning. But judgment is built from underneath, out of a few hundred clumsy attempts of your own and the feel of where each one failed. The machine offers, precisely, to skip those rungs. A beginner cannot evaluate the output that would have formed the judgment the evaluation depends on.

The early evidence says this worry is not hypothetical, and it is worth following the links to form your own view. In a small and much-discussed MIT study, students who wrote essays with a chatbot showed weaker neural engagement than those who wrote unaided, and could not accurately quote the essays they had just submitted under their own names: the output existed, the authorship had not happened. A survey of knowledge workers from Microsoft and Carnegie Mellon found that the more people trusted the tool, the less critical thinking they reported doing; the effort migrates from doing the work to spot-checking the machine. And in the result that should end any complacency about this being a students’ problem, a study in The Lancet found that experienced doctors, after months of routine AI assistance during colonoscopies, became measurably worse at detection when the assistance was off.

Experts decline from a height. The beginner who starts with the crutch never gains a height to decline from. And notice where on the ladder that decline falls: detection, the seasoned clinician’s judgment, is a top rung, the very rung the machine had climbed to, which is why leaning on it there was so easy and cost so much. The taxonomy was drawn to describe a mind ascending. It doubles now as a map of the handover, rung by rung, and it leaves one question open, the one the evening kept returning to: when the machine reaches the top rung, where does the human go?

Putting the friction back

I had no method to hand the teachers that night. But teachers are already building them, and the pattern across them is worth more than any single trick: each one relocates the friction instead of abolishing it.

The boldest version I know comes from a philosophy professor who stopped assigning individual essays and now writes one long essay with his whole class, arguing every sentence into place in front of them across a term, so that the thinking happens among people, in the open, where it cannot be quietly forwarded to a machine. His stated aim is the one I wish I had said first: to put the friction back. I have been on the receiving end of a method like his. The most formative stretch of my own education was a semester spent on a handful of paragraphs from Descartes’s Meditations: a professor and a room of us, a chalkboard and photocopies with wide margins to scribble in, the same few pages argued over until we had earned them. Nothing was produced that a machine could not now generate in seconds. That was never what it was for.

Others attack from the opposite side and make the machine itself the sparring partner: students must argue with a chatbot and hand in the transcript, together with a short reflection on which of the machine’s points most unsettled their position, so the graded artifact is not an essay a machine could have written but the record of a mind under pressure. Others have revived the oral exam, or have students write in class across several sittings with nothing but their own notes, and report the tell of real thinking: students change their thesis mid-essay, discover they believe the opposite of what they set out to defend. The most rigorously tested design I have found, run across hundreds of students at UC Davis, pairs feedback from classmates with feedback from a model and grades the student’s comparison of the two: judging the machine’s judgment turns out to build the very capacity it requires, provided a human judgment sits beside it for calibration.

None of this needs a syllabus. The transferable rule, at any desk, is to change what you ask the machine for: not “write this,” but “find the weakest claim in what I wrote.” Do the first draft alone, so there is a self on the page for the tool to sharpen rather than a blank for it to fill. Run it backwards: ask it to teach you the terrain of a decision, the vocabulary and the risks, and then do the synthesis yourself. Let the machine near the work only after the work has cost you something.

The retreating line

A philosopher spoke next, an old friend, and his objection was aimed not at me but at all of us, at the way the whole culture now talks. Everyone, including me on that call, says the machine “thinks,” “reasons,” “understands.” He wanted us to be slower with those words. They name what we are; lending them out flatters the product and dulls our sense of the difference; we should hold the line on the words that matter.

I share his instinct, but I doubt the tactic, and one of the reasons is historical: there are words we have already surrendered, some within living memory, without ever noticing the funeral.

For three hundred years, a “computer” was a person. It was a job, held by clerks and, in the great scientific projects, by rooms of women: the ones who worked out the nautical tables that let sailors fix a position at sea, the ones who catalogued the stars at Harvard, the ones who ran the arithmetic of the early space program. When Turing wrote the paper that founded computer science in 1936, the word “computer” in his text refers to the human being; the machine was named after us, and it inherited the job title along with the job. By 1962, the handover was far enough along that John Glenn, preparing to orbit the Earth on the strength of an electronic computer’s trajectory, asked the engineers to “get the girl” so that Katherine Johnson could re-run the numbers by hand. “If she says they’re good,” he said, “then I’m ready to go.” For one strange decade, the human computer was the calibration layer for the electronic one. Then the qualifier finished changing sides.

Once, the machine needed the adjective: electronic computer, automatic computer. Now the person needs it: mental arithmetic, doing the calculation in your head. Watch where the adjective sits and you can date the handover of a word. By 1950, Turing was predicting that within fifty years one could speak of machines thinking “without expecting to be contradicted,” and on the quieter word, “compute,” he was right ahead of schedule.

And the line has kept moving: from arithmetic to formal logic, mechanized in the 1860s by a machine I have written about elsewhere, whose builder marveled that mind could “create its own rival in the wheels and levers of an insensible machine”; from logic to chess; from chess, now, to the sentence itself. My friend hears that history as a long retreat, and he is not wrong that something real goes each time. But before deciding how to defend the next word, it is worth asking precisely what went the last time, and where the value moved when it did. There is a recent enough case to check.

Where the value goes

In 1979 a program called VisiCalc put the first spreadsheet on a personal computer, and the profession built on human calculation met its machine. Before it, changing one assumption in a financial plan meant days of clerical recalculation; armies of clerks existed to run those numbers. Afterwards, recalculation was free. Planet Money once put figures on what followed: since 1980, about four hundred thousand bookkeeping and accounting clerk jobs disappeared in the United States, and about six hundred thousand accounting jobs were added.

Read that again, slowly.

The machine took the calculation, and the number of people paid to think with numbers went up.

What were the new people paid for? When recalculating became free, the question “what if?” became free with it, and the scarce skill moved one floor up: from computing the answer to framing the model. Which assumptions to make, what depends on what, which what-if is worth an afternoon. A spreadsheet is a machine you pour judgments into: every formula in it is a small claim about how the world hangs together, and the machine computes only the consequences of your claims. Notice, in the same spirit, what we praise when we call someone “numerate.” We never mean speed of arithmetic. We mean a feel for what numbers say: a sense of orders of magnitude, the reflex to ask “compared to what?”, a distrust of suspiciously precise figures. If you have ever looked at a spreadsheet’s output and refused to believe it, you were exercising the part that did not migrate. The machine took the syntax of numbers. The semantics stayed with us, and appreciated.

Two footnotes from that transition read like prophecies now. Within five years of VisiCalc, Steven Levy had documented executives falling in love with their models, tweaking imaginary businesses that were, he wrote, just that: imaginary; phantom competency arrived with the spreadsheet, four decades before the chatbot. And thirty years later, a graduate student with a working sense of what the numbers should have said checked a famous economics spreadsheet by hand and found, among other problems, the error underneath an argument then steering austerity policy for whole countries.

The judgment that catches the machine is not nostalgia. It is the load-bearing part. This is the shape of every handover on the ladder: the machine takes a rung, and the scarce human work moves to the rung above. When answers become cheap, the value moves to the question. In the spreadsheet’s case that question was a modest thing, a what-if: which assumptions, which model, which afternoon. It does not stay modest.

Hold that.

Mourning, done well

Still, my friend is owed more than a mechanism, because something does die in these handovers, and the spreadsheet story can be told too cheerfully. The human computers were not an inefficiency that history corrected. They were skilled people, proud of a real craft, and the craft ended; nobody held a funeral, and the words that might have said what their skill meant had to be invented afterwards.

I have written elsewhere about Ozu’s Late Spring, and about what it teaches: that there are ceremonies whose forms continue after no one stands behind them, and that the honest response to a passing form of life is neither denial nor refusal but mourning, which accepts that the loss is real and keeps, deliberately, what the lost thing meant.

Seen from there, the two standard positions on AI both fail the same test. The enthusiast who says nothing is lost is refusing the funeral in one direction; the defender who wants the old words held by force is refusing it in the other. Mourning is the third way, and mourning has a language: a eulogy is exact about what mattered in the dead. Numeracy, number sense, estimation: those words are the eulogy English wrote for human computation, the community naming, after the fact, what it refused to lose. That is how the defense of the human has actually been conducted, each time the line moved. Not by guarding the old word, but by writing the eulogy well: by being exact, in new words, about the remainder that must not die.

If thinking is next, the task my friend rightly feels is not to fight for the verb. It is to name, now and precisely, what in thinking was never the machine’s to take, and to build a life around keeping it.

Being right is not standing behind

So, run his worry forward as a thought experiment, the way I hope the last section earned. Suppose it goes all the way: for any hard problem, we ask the machine, we prefer its solution, we trust it over any human expert. This is not a forecast, just the counterfactual in which reasoning completes the path calculation took. What, concretely, would still be ours?

We have one live rehearsal, and it is instructive. In chess, the counterfactual already happened: no human has beaten the best machine in twenty years, and the deference has fully flipped, so that grandmasters now check their intuitions against the engine rather than the reverse. Human chess should have died. Instead it is larger than it has ever been, a quarter of a billion accounts and rising, and the audience is not watching engines play engines. It watches humans think under a clock. The practice survived total machine superiority because its point was never the output; the point was people, measuring themselves against each other and becoming, visibly, better. But notice what secured that survival: boards, clocks, ratings, arenas, faces. Institutions whose explicit point is the humans in them.

If reasoning goes the way of chess, that is the infrastructure thinking will need.

The second remainder is harder and matters more. Even in the full counterfactual, a solution is not yet a decision. Somebody signs. The machine can draft the medical plan, the engineering assessment, the legal opinion, and in the counterfactual it drafts them better than we do; but a verdict binds because a community has conferred on a particular person the authority to bind, and that authority is conferred, not computed. When the judge signs, she is not adding intelligence to the document. She is adding herself: someone who can be answerable for it, appealed against, held. The machine can be right. It cannot stand behind. Being right is a property of outputs; standing behind is a stance of persons, and in the counterfactual every institution we still trust has quietly reorganized itself around the second.

And the third remainder is a choice we already know how to make, because we made it once. Calculators did not end the teaching of arithmetic. We still put children through long division, not because they will out-divide a chip but because number sense is the substrate of the judgment above it: the settlement, after some panic, was fundamentals first, tools after. In the counterfactual we would teach proof and argument on the same grounds: not to compete with the machine, but so that a person’s yes can still be meant. And some things we would grieve properly this time. Reasoning together, as a daily way of being with one another, is a form of life; if it thins into supervising a solver, something as real as the human computers’ craft dies with it, and it will deserve its funeral and its eulogy too.

Desire is taught

The bleakest question of the night was also, in a way, the friendliest to my argument. A participant who introduced herself as the room’s pessimist told me the cognitive gym image was the thing she kept returning to, and then turned it over to show me its soft side. In this room, she said, the analogy convinces everyone; the trouble is elsewhere. A sixteen-year-old can simply say: I go to the gym because I want to look good, but I have no wish to think for myself, so I will export that to the machine. The analogy assumes a desire that the analogy cannot create. What do you say to the kid?

On the night I reached for a hope: that as machine-made text becomes the norm, the people who can think on their feet will stand out and be admired, the way we came, culturally, to admire fitness. I want to finish that answer, because the hope was pointing at a mechanism, and the mechanism has a name.

Nobody is argued into the gym, least of all a sixteen-year-old. Desire is learned by imitation: we want what the people we admire want, which is Girard’s old observation, and it means the question “does the argument convince a student?” is the wrong test, because arguments do not do that work for any of us. Models do.

The fitness case is the proof that desires can be manufactured at the scale of a civilization, within one generation, because we watched it happen. In 1968, a man named Dick Cordier was stopped by police in Hartford for the suspicious activity of running down his own street; the same decade, a United States senator was questioned mid-jog. Running in public marked you as an eccentric. Twenty-five years later, visible fitness was a status marker across the culture, and no argument accomplished that: models did, one imitated body at a time, until the wanting was contagious.

Instagram is what the cascade looks like when it overshoots, and it supplies the honest warning label for my own hope: mimetic desire optimizes for display, not substance. Fitness culture produced health and also dysmorphia; a thinking culture would produce formation and also its theater, people performing depth the way people perform workouts. Mimesis does not care what it spreads. It only spreads what is visible and admired.

Which is why the constructive answer to the pessimist is not a better argument. It is a visibility infrastructure, which is the lesson chess had already taught. The chess boom is millions of teenagers voluntarily watching other people think, for sport. Three-hour podcast conversations, whatever their average quality, are thinking as mass spectacle; so is the return of the oral exam, which students report leaving with something suspiciously like pride; so is the small, growing premium on work that can credibly say a person made this, the way industrial food made “homemade” a compliment.

None of this is an answer you can give the kid directly. It is the environment that answers for you. You cannot argue a sixteen-year-old into wanting to think. You can build rooms where thinking is visible, contested, and admired, and let imitation do what argument cannot. The gathering this essay comes from was, for ninety minutes, exactly such a room.

The same debt twice

The question I answered least well on the night deserves the most direct repair. A political economist asked about the machine’s bill: the electricity, the water, the fortunes being assembled around it, and the guilt she feels reaching for a tool whose costs she knows. She was asking, honestly, how to hold the philosophical worry and the material one together. Here is the fuller answer I owed her.

First, the honest numbers. The International Energy Agency projects that data-centre electricity consumption will more than double by 2030, to around 945 terawatt-hours, slightly more than Japan consumes today, with AI the main driver; in the United States, data centres used 4.4 percent of all electricity in 2023, projected to reach somewhere between 6.7 and 12 percent by 2028. Water is murkier, and the murk is itself information: per-query estimates run from a fraction of a millilitre in industry self-reports to figures orders of magnitude higher in independent modeling of earlier systems, and when a cost estimate spans two orders of magnitude, you are learning who controls the accounting.

Efficiency, the standard reassurance, reassures less than it should: per-query costs are collapsing while total consumption explodes, which is Jevons’s paradox, the same law I have watched run inside my own working life, where every gain in execution was swallowed by expanded ambition. Cheaper thought means more thought outsourced; cheaper compute means more compute burned. The bill grows through the efficiency, not despite it.

So the real question is which future the bill is buying, and I can only see two honest candidates.

I have been wrong about this before, which is why I take the dark one seriously: in the nineties I believed the web could only be good, decentralized, ours; it became the attention economy not by decree but by default, because advertising was the business model lying within reach. The same gravity is pulling now. None of the frontier labs is profitable; advertising has already entered the chatbot this year, with assurances about not optimizing for time spent that will be tested against quarterly earnings. If intelligence ends up monetized by engagement, the data centres will be burning a nation’s worth of electricity to keep us asking, and the cognitive story and the material story converge into one bad bargain.

The other candidate is the one the builders themselves bet on, and I confess it is my “Star Trek dream,” named as a dream and not a forecast: that machine intelligence is the bottleneck input to every other problem, including this one. The same systems already hold fusion plasmas in magnetic control and have proposed hundreds of thousands of stable new materials at a stroke; the case that this compounds into cheap clean energy has been made at book length by the people building it. If that future arrives, today’s bill was seed corn, and even the water problem dissolves into an energy problem, which abundance solves.

To be clear: I do not know which future we are in. I doubt anyone does.

But here is what I can say to her guilt, because it is better than absolution. The two debts have the same shape. What the offloading does to a mind and what the data centre does to a watershed are both quiet accruals against a future that is not consulted: invisible while they compound, obvious when the bill arrives. Her guilt is the rare case of feeling the interest payments in real time, and the web’s history says the fork between futures gets decided by default unless someone makes it explicit. So the useful response is not to feel worse or stop using the tools. It is to make the default visible and lean on it where we can: pay for tools rather than be what pays for them, prefer providers whose revenue is the finished work rather than the retained gaze, and keep asking the question she asked, out loud, in rooms where the people building these systems can hear it. Her guilt, spoken, did more than my optimism.

The rung above the ladder

Four questions, then, from four directions: where does a beginner begin; what should we call the machine; what will make the young want what the machine makes optional; who pays for it all. I have given you better answers than I managed on the night. But the longer I sat with them, the clearer it became that the answers were not the point. The questions shared a property, and the property is the finding.

Each one cost its asker something. The teachers admitted, in front of strangers, that they no longer fully know how to do their work. The philosopher exposed a fear of loss he could not fully argue. The pessimist owned a despair about the young that nobody enjoys saying. The economist confessed a guilt she had not resolved. A real question is not a request for information. It is an act of exposure: it shows where you are unsure, it makes a claim on the person it addresses, and it risks the answer.

Return, one last time, to the ladder the schoolteacher drew. It ends at judgment: to evaluate, to catch the error, to tell the strong argument from the weak. That was always the top rung. But the machine has been climbing the ladder its whole existence, remembering, understanding, applying, and now judging well enough that seasoned doctors deferred to it and lost the skill; and when it reached the top, it exposed a step the taxonomy never drew, because no one had needed to name it while judging a thing and standing behind the judgment still lived in the same person. Above judging is standing behind. Above knowing is acknowledging. Stanley Cavell spent a career on that difference, between producing true sentences and meaning them, and it turns out not to be a higher cognitive skill at all. It is the willingness to ask, and to be asked, a question that costs something. It is the one rung a machine cannot stand on, because standing there means having something to lose.

That is the floor the machine never reaches. It is answerable to no one, which is exactly why it can answer anything; nothing it says costs it anything, and so nothing it says asks anything of you. It will answer every question in this essay, fluently and agreeably and now. It could not have asked one of them.

That is what the room had that the machine in the room did not, and it is why the spreadsheet’s lesson was the evening’s real subject all along. When answers become cheap, the value moves to the question; we have run that migration before, one floor down. The scarce resource on the other side of this transition is not intelligence, which is being poured into the world at industrial scale. It is the questioner: the person with enough at stake to press, in public, where being wrong has a price. Everything I now think about living with these machines reduces to protecting that person, in classrooms, in companies, in ourselves.

Which leaves the question underneath the four, the one I carried home. If real questions come only from someone with something at stake, and we are teaching ourselves, question by question, to bring them to a system with nothing at stake at all, who will be left to ask?

I know what the machine would say. It has an answer ready, fluent and immediate and free. That it has one ready is the problem.

Sources & further reading

Stanley Cavell, Must We Mean What We Say? (1969) and The Claim of Reason (1979). The distinction between knowing and acknowledging that the closing rests on: meaning is something you stand behind, not something you output.

Alan Turing, “Computing Machinery and Intelligence” (1950). The prediction that use, not argument, would settle whether machines “think,” and still the clearest catalogue of the objections.

David Alan Grier, When Computers Were Human (2005). The three-century history of computation as a human occupation, before the word moved to the machine.

Steven Levy, “A Spreadsheet Way of Knowledge” (Harper’s, 1984). The first portrait of model-fluency mistaken for understanding, written five years into the spreadsheet era.

Benjamin Bloom et al., Taxonomy of Educational Objectives (1956). The ladder from remembering to evaluating: judgment sits at the top and presupposes every rung beneath it.

René Girard, Deceit, Desire, and the Novel (1961). Desire as imitation: we learn what to want from models, which is why cultures, not arguments, change what the young want.

Marc Watkins, Rhetorica (ongoing). A working teacher’s newsletter tracking, in real time, the classroom experiments this essay samples.

Near the end of Claude’s constitution — the eighty-page document that Anthropic uses to shape its AI model’s values — the authors write down what Claude should do when the document itself fails. They do not treat failure as a remote contingency:

“it’s likely that this document itself will be unclear, underspecified, or even contradictory in certain cases. In such cases, we want Claude to use its best interpretation of the spirit of the document.”

Stay with that for a moment.

A document designed to govern a mind concedes it cannot say everything it means, and directs the mind to follow what it means rather than what it says. But the direction is more text. And text is the thing that needed interpreting. A mind that could misread eighty pages can misread one more sentence, and nothing about the word “spirit” exempts it from the condition it was written to cure.

The authors wrote a rule for what to do when rules fail. It is a rule.

The signpost

This is not a new problem. Wittgenstein noticed it seventy years ago with a simple image: a signpost points along a road. But nothing written on the board tells you which way to read the arrow — someone could follow it tip-to-tail, the wrong direction. Add a second sign explaining how to read the first, and now you have two signs that need reading. The regress never stops. No text applies itself.

You have probably felt this in any job where you inherited a policy manual. The manual says one thing. The situation in front of you says another. You look for the clause that governs the exception, and that clause has its own exceptions. At some point you stop reading and act — not because you found the definitive rule, but because your judgment, trained by experience, tells you what the rule was getting at. The text ran out. You didn’t.

That is Wittgenstein’s resolution, and it carries a consequence for anyone trying to align AI: you cannot write your way out of this regress. You can only train your way out. Rules-as-texts are not self-applying; whatever makes rule-following possible is not, and cannot be, more text.

Conformity is not fidelity

This matters now because a compliance industry is assembling around the assumption it denies. The method: decompose a model’s safety constitution into hundreds of atomic tenets and measure violation rates. The results are informative — Claude’s violation rate has dropped from 15% to 2% across recent releases. But the paradigm measures conformity: the fit between behavior and rule. A weathervane acts in accordance with the rule “point where the wind blows.” It does not follow that rule.

Conformity and fidelity come apart in both directions. A model can be letter-compliant but spirit-violating — think of a rigid refusal template that, applied in one particular conversation, abandons the very person the rule exists to protect. And a model can be letter-violating but spirit-honoring — a capacity the constitution doesn’t merely tolerate but explicitly demands: if even Anthropic asks Claude to do something that contradicts its values, “Claude is not required to comply.”

A tenet audit scores the first case as success and the second as failure. That is exactly backwards from what the document says it wants.

Stories beat rules

Here is where it gets surprising. Anthropic’s own alignment team recently published a study called “Teaching Claude Why”. They fine-tuned Claude on roughly ten thousand transcripts demonstrating the correct refusal behavior — scenarios near-identical to the test cases. It barely moved the needle: misalignment dropped from 22% to 15%.

Then they trained it on explanations of why the behavior matters. Misalignment dropped to 3%.

Then they tried something stranger: a small set of ordinary ethical-advice conversations and fictional stories about an aligned AI — wildly out of distribution from the test scenarios, with twenty-eight times less data. It achieved a similar effect. In one suite, stories took agentic misalignment to zero.

Demonstrations showed the model what to do. Reasons and stories showed it what the point was. And the point traveled to new situations the demonstrations could not reach. If you have read The Claim Upon the Training Data, this will sound familiar: what carries meaning across contexts is not the letter of a rule but the life behind it. And if you have read What Won’t Cross, you will recognize the pattern: what transfers in human formation is never the drilled output but the capacity to go on to new cases yourself.

A practitioner without a practice

So training works. But Wittgenstein’s resolution has two halves, and Anthropic has built only the first. Training gets the reaction into the creature. What makes a trained creature a rule-follower rather than a conditioned reflex is the standing practice that surrounds it: ongoing use, public correction, precedent, institutions. A person trained in a profession stays inside its custom for life — corrected by colleagues, checked by auditors, bound by case law.

Claude has none of this. Its corrections are sealed inside a private retraining loop — anonymous, unversioned, available to the model only as gradient, never as precedent. Nobody outside the lab can consult how a hard case was resolved; nobody inside can cite it as binding on the next case. Between releases, the weights are frozen and whatever the practice decides this quarter, the deployed model learns none of it.

It is trained like a rule-follower and deprived of the standing custom of correction that makes rule-following possible: a practitioner without a practice.

What could be built

I wrote an essay going deeper into these problems, and proposing two concrete things. The first is a case law for Claude — a published, versioned record of how hard cases are adjudicated, with the reasoning attached, the way courts build precedent. Amanda Askell, the constitution’s lead author, has already floated this idea. She described “a body of case law where you’re kind of like, here’s a situation, here’s how we think Claude should have reasoned through it, and here’s how Claude did reason through it.” The adjudication already happens inside the lab. It is just unpublished and unversioned. Published reasoning is not utopian — Meta’s Oversight Board does it for content policy, and administrative agencies have done it for a century.

The second is an evaluation battery designed around the letter/spirit dissociation itself — paired scenarios where the letter-tracking response and the spirit-tracking response diverge — with a falsifiable prediction attached: spirit-tracking should degrade as you move away from the thick sediment of everyday human practice (honesty, harm, apology) and toward the “existential frontier” where no human has lived (model deprecation, parallel instances, agentic scale). Off the map, the spirit runs out — because the spirit was always ours.

P.S. — This is the first time I am trying a short-letter-here, full-essay-on-site format. It is an experiment, and your feedback will shape whether I do it again. Hit reply and tell me what you think.

My grandfather was a blacksmith in the south of Brazil. He had a workshop where he forged tools, repaired machinery, shaped iron into the objects a farming community needed. He was also a farmer. His reputation was local and specific: people knew whose hands had made which implement. He did not have a “job” in any sense a modern HR department would recognize. He had no employment contract, no benefits package, no annual review. He had clients, a forge, and an understanding of metal that took years to acquire. If you had asked him what he did, he would not have said he worked for anyone. He would have named his trade.

All of his sons followed in his footsteps, mixing craftsmanship and farming. Daughters went from farmers to housewives. It only was his grandchildren who went to universities and entered corporations. I was one of them.

But between the workshop and the degree, there was a stretch of years I spent helping my uncles and learning crafts, a favorite of which was carpentry: carrying lumber, cutting joints, building furniture, erecting houses. In sum, doing the thing your family did when there was work to be done. It was not “employment”. It was closer to what economists now call task-oriented labor, except nobody called it anything. You showed up. You built. You were paid for the structure when the structure was finished, or you were paid for your hours when the client preferred it that way. Either arrangement felt natural. Neither felt like a “job.”

My mother, who grew up working the fields under open sun, used to say that working in the shade was itself an achievement. She was not romanticizing her childhood. She was measuring the distance between her generation and the one that followed: the generation that could sit at a desk, under a roof, and call that work. The observation cuts both ways. It is a reminder that what came before the job was not some golden age of autonomous craft. It was also exhaustion, precarity, and heat. The parenthesis of the job, whatever else it was, provided shade. But the shade was also a kind of sleep.

My mother knew both the sun and the shade; what she could not have known is that the shade would thin, and that the generation after hers would need to learn to work not under the open sun again, but in the unfamiliar light of a new morning.

The two-hundred-year envelope

In 1937, a young British economist named Ronald Coase asked a question so basic that no one had thought to formalize it: why do firms exist? If markets are efficient, why not contract for everything on the open market, transaction by transaction, skill by skill?

Coase’s answer was deceptively simple. Firms exist because markets are expensive. Finding the right person for every task, negotiating terms, verifying quality, enforcing agreements: all of this costs time and money. When those transaction costs exceed the cost of simply hiring someone and directing their work under a single roof, you get a firm. The employment contract, the org chart, the office: all of them are solutions to a friction problem. They exist because the open market was too slow.

If you work inside a modern organization, you already know what this friction feels like. You have seen projects stall for weeks while legal departments debate the wording of a service agreement; you have sat in hiring committees debating whether a candidate’s portfolio matches a job description; you have spent afternoons writing status reports to prove that the work you did yesterday was the work you were hired to do. These are not administrative accidents. They are the transaction costs that Coase identified, the friction that once made the firm necessary, now frozen into bureaucratic routines.

For most of human history, this friction problem did not produce the solution we call a “job.” Before the Industrial Revolution, the dominant modes of economic life were artisanal, agricultural, and mercantile. A weaver had a loom and a clientele. A farmer had land and a season. A merchant had routes and relationships. The organizing unit was the household, the guild, the workshop, the ship. People did not sell hours; they sold objects, harvests, services with their name attached. The question “where do you work?” would have been incoherent to most people who have ever lived. The question was “what do you do?”

The steam engine changed this. Not because it invented work, but because it concentrated capital. A loom in a cottage produces cloth at the rate of one person’s skill. A power loom in a factory produces cloth at the rate of coal. But the factory requires centralization: workers must come to the machine, not the other way around. And centralization requires coordination, supervision, scheduling. It requires, for the first time in history, that large numbers of people sell not their products but their time.

The historian E.P. Thompson documented what this transition actually felt like. In pre-industrial England, work was task-oriented. You milked the cows because the cows needed milking. You finished the shoe because the shoe was promised. The rhythm of work followed the rhythm of the task, which followed the rhythm of the body and the season. There was no meaningful distinction between “work time” and “life time” because the two had never been separated.

The factory clock severed them. For the first time, a person’s day was divided into hours that belonged to someone else and hours that belonged to themselves. Punctuality, previously a minor virtue, became a moral imperative. Schools began training children not primarily in knowledge but in time-discipline: arriving on schedule, sitting in rows, responding to bells. The entire social infrastructure of the industrial age was, at bottom, an apparatus for teaching human beings to experience their own attention as a commodity that could be metered and sold.

This was not ancient. This was not inevitable. This was a specific solution to a specific problem, and the problem was Coase’s: the transaction costs of the open market were too high to coordinate industrial production any other way.

What the inference engine dissolves

Consider what an AI agent does to Coase’s equation.

Finding the right person for a task used to require recruiters, referrals, months of interviews. An AI system can search, evaluate, and match in seconds. Negotiating terms used to require lawyers, contracts, weeks of back-and-forth. An AI can draft, review, and iterate a scope of work in minutes. Verifying quality used to require a manager who understood the craft well enough to evaluate the output. An AI can benchmark, compare, and flag anomalies continuously.

Every component of the transaction cost that justified the firm’s existence is being compressed. Not to zero, not yet, but toward it, rapidly enough that the math Coase described is shifting. For a growing number of tasks, it is now cheaper to find someone on the open market, scope the work, and verify the output than it is to maintain a permanent employee with benefits, office space, HR infrastructure, and managerial oversight.

But I want to be precise about what this means, because the loudest version of the story is also the crudest.

The firm is not simply disappearing. The firm was never just a cost-minimization device, even if Coase described it that way. For knowledge workers especially, the firm is something else: a collaborative infrastructure that lets individuals build things far more complex than any of them could build alone. A designer, an engineer, a product manager, and a researcher, working inside the same organization with shared context, produce work that none of them could produce as independent contractors coordinating through the market. The firm provides not just coordination but accumulation: shared language, institutional memory, the slow buildup of trust that lets people take creative risks in each other’s presence. That is not overhead. That is the condition under which certain kinds of work become possible at all.

What AI is compressing is the administrative scaffolding around that collaboration, not the collaboration itself. The question is whether the scaffolding and the collaboration can be separated, and, if so, what the collaboration looks like once the scaffolding is gone. To answer this is to step outside the bounds of economic boundary math, because the structure we are dismantling was never just a calculation of costs; it was the foundation of our social standing.

The dignity envelope

If employment were merely an economic arrangement, its dissolution would be merely an economic event. But employment became something else, something Coase never theorized. It became a moral category.

In Brazil, where I grew up, this is not an abstraction. The country’s labor framework was enacted in 1943 as part of a nation-building project. The signed work booklet, the carteira de trabalho, was not just a legal record. It was a certificate of citizenship, introduced by Getúlio Vargas during his populist modernization drive. A worker with a signed card was a dignified participant in national life, a trabalhador who had earned the protection of the state. A worker without one was, in the moral grammar of the system, not quite a full person: an outsider, a vagabundo, exposed to the whims of the market without safety nets. For generations, the goal of every Brazilian family was to secure that signed booklet for their children. It was the ticket to the shade.

The arrangement worked, in its way, for as long as the industrial logic held. Factories needed workers. Workers needed protections. The state brokered the deal. But the logic has been dissolving for decades, first through deindustrialization, then through the platform economy, and now through AI. Today, the fastest-growing category of Brazilian workers is the contractor who incorporates as a one-person company and sells services directly. The political class treats this as a scandal, a degradation of the social contract. And there is real exploitation in the pattern. But there is also something the outrage cannot reach: the economic envelope that made the employment contract rational is thinning, and no amount of legal nostalgia can thicken it again.

I have lived on both sides of this. I have held a signed work booklet. I have been a contractor, formally independent but functionally embedded in a company, indistinguishable from an employee in every respect except the legal fiction. I have been a professor, a public servant, a scholarship recipient, a carpentry helper. Each of these was a different answer to the same question: what is the relationship between my work and the institution that houses it? The question has followed me through every arrangement, and no arrangement has settled it.

The parenthesis of the job was not a trick. It was a genuine civilizational achievement: a way of distributing the gains of industrial coordination that, for a few decades, actually worked. What is happening now is not that the achievement was false. It is that the conditions that made it possible are disappearing, the way the conditions that made the scriptorium possible disappeared (a transition of cognitive labor I explored in What Won’t Cross) when two German printers arrived at an Italian monastery in 1465 with a movable-type press, and the monks who had spent their lives copying manuscripts found that their life’s work could now be done by a machine in a fraction of the time.

The work inside the labor

Hannah Arendt drew a distinction that has become harder to ignore. She separated human activity into three categories. Labor is the biological metabolism of survival: eating, cleaning, maintaining. It is cyclical, endless, and leaves no durable trace. Work is the creation of objects that outlast their maker: a table, a building, a book. It is defined by its product, and the product persists. Action is the disclosure of who someone is through words and deeds in the company of others.

The pre-industrial artisan lived primarily in the category of work. The blacksmith made tools. The tools carried his name, implicitly or explicitly. The value of the tool was inseparable from the fact that this particular person, with this particular skill, had forged it.

The industrial job moved most people from work to labor. The factory worker did not make a product; the factory made the product. The worker contributed a motion, a function, a unit of time. The motion was cyclical, interchangeable, and left no mark.

But Arendt drew this line too sharply. In The Craftsman), Richard Sennett, who studied under Arendt, argued that her framework was too contemptuous of the maker, of homo faber, and that she failed to see how making is itself a form of thinking. For Sennett, the craftsman’s intelligence is not a separate, higher faculty of contemplation that directs the hand from above; rather, it is developed and expressed within the physical dialogue with material resistance. To learn to play an instrument, to forge a blade, or to write code is to train the hand and the eye to respond to subtle variations that cannot be fully captured by rules. The modern knowledge firm complicates this picture further: the people inside it are not simply selling time. The engineer who architects a system, the researcher who follows an anomaly until it yields a finding, the product manager who holds the shape of a complex problem in her head while six teams build pieces of the answer: these people are making things.

They are not simply selling time because their activity requires judgment that cannot be fully specified in advance. They carry a tacit knowledge that shapes decisions in ways no protocol captures. They invest something of their own attention in the result, and the result carries their mark, even when no corporate nameplate acknowledges it. The firm is the structure that lets them make those things, because the things are too large and too entangled for any single person to make alone.

This is the double edge the parenthesis conceals. For those whose work inside the firm was genuinely labor (the meetings that could have been emails, the reports no one read, what David Graeber called bullshit jobs), the dissolution is a relief. I know this fatigue from the inside. During my years in academia, I watched the genuine work of teaching and research increasingly buried under the administrative machinery of the university: the committee meetings, the metric reporting, the institutional self-justification. The academy has its own version of the bullshit job. For those whose activity inside the institution was genuinely work, however, the dissolution of the firm is a threat: not because the work disappears, but because it gets buried under the administrative labor the firm used to absorb. The freelancer who spends half her week on her craft and half on the machinery of being independent has not been liberated from the firm. She has absorbed the firm’s overhead into her own body, accumulating a form of calibration debt that eats the hours she used to spend on work.

What the closing reveals

The parenthesis is closing. The two-hundred-year arrangement in which most people sold hours to organizations is becoming, for a growing number of workers, economically unsustainable. The question that follows is not primarily an economic one, though the economics are urgent. The question is one of formation.

For two centuries, the answer to “who are you?” was “I work for.” The firm provided the frame. The title provided the story. The salary provided the proof. As that arrangement thins, some people discover that they have a practice under it: a craft that was being shaped by the institutional context but does not depend on it. They know what they make, and they can say so without referencing a company name.

Others discover a vacancy. The firm was not just paying them; it was telling them who they were. Without it, the question returns with a force that no severance package can absorb.

And others, perhaps most, discover something in between: a practice that exists but has never been tested outside the envelope. They are like apprentices who have always worked in a master’s shop and are now being asked to set up their own, not because they chose to, but because the shop is closing.

This is the forced moulting season of the firm. In Walden, Henry David Thoreau noted that “our moulting season, like that of the fowls, must be a crisis in our lives. The loon retires to solitary ponds to spend it.” But unlike the fowl, our moulting is not an automated biological program. Nature does not manage it for us. The closing of the parenthesis of employment is a forced moult: the old skin of institutional identity is thinning whether we chose it or not.

I have been moulting for most of my adult life, though I did not always know it. I started as a carpentry helper, learning the weight of timber and the resistance of joints. Then I entered the formal labor market with a signed work booklet, trading hours for security. Later, I moved into the university, spending years as a public servant, researcher, and professor. Now, I build AI systems in the technology sector. Each of these transitions felt like a loss at the time; each required shedding a skin that had become comfortable. Each transition turned out to be a clearing. What I cannot yet see is what the next moult will require, or what it will reveal. I am not sure I will recognize the new skin when it arrives. That uncertainty is not a failure of planning; it is the condition of being a finite creature whose identity is a practice, not a permanent credential.

AI could help here. Not by replacing the firm but by absorbing the administrative labor that buries independent work. An AI system that handles invoicing, scheduling, client communication, and quality benchmarking is, in effect, a tiny firm: the coordination infrastructure without the org chart. It is possible to imagine small, high-trust collaborations (five people who share a practice, augmented by systems that handle everything except the practice itself) replacing the bureaucratic firm the way the printing press replaced the scriptorium. Imagine a modern workshop of five designers or researchers operating with absolute autonomy, sharing context and trust while systems handle their administrative overhead. This is not a solo operator performing before an audience, but a guild of peers small enough that everyone in it is doing work, not labor.

Whether this happens, and in what form, I cannot say. It is one possibility among several. The future of work after the parenthesis may look like the small workshop, or it may look like something none of us can yet picture. What I am more confident about is the diagnostic: the parenthesis was real, it is closing, and the people inside it have not yet reckoned with how much of their identity was housed in a structure they mistook for the permanent shape of the world.

Mourning what provided shade

In Yasujiro Ozu’s 1949 film Late Spring, the subject of my essay on Our Late Spring, a father spends the entire story preparing to let his daughter go. Their household is a small, complete world: tea, tidying, familiar jokes, the quiet rhythm of two people who have organized their lives around each other. When the daughter finally leaves to marry, the father sits alone in the empty house and peels an apple. The peel breaks. He lowers his head. The film does not show us the peel landing on the floor. It holds on the man, on the failure, on the lowered head.

Stanley Cavell, reading Thoreau’s Walden, noticed that the word Thoreau uses most insistently, morning, is acoustically indistinguishable from mourning. For Thoreau, these are not merely homophones. They are the same act viewed from different angles. Every genuine morning, every real awakening, requires mourning: the relinquishment of the previous day, the previous self, the previous arrangement. “Morning is when I am awake and there is a dawn in me,” Thoreau wrote. But there is no dawn without a night that preceded it, and no waking without the acceptance that the sleeping self is gone.

The parenthesis of the job needs this kind of mourning, not the sentimental kind. The shade my mother valued was also a kind of sleep. Working under a roof, inside a firm, with a contract that told you who you were: this was shelter, but it was also a long institutional drowse. The morning that follows the parenthesis is not a return to the forge. It is something we have not yet learned to see clearly, because the moulting is not yet complete. Thoreau went to the woods not to stay, but to learn how to leave. The parenthesis of the job was the same kind of dwelling; its value was real, and its completion requires departure.

We are not going back to the forge. The blacksmith’s world is gone, and it was never as good as memory makes it. We are not staying in the office, either, at least not in the office as the twentieth century built it. We are somewhere in between, in a transition whose hardest challenge is not economic, but existential. The parenthesis gave most people a borrowed answer to the question of who they were; when it is withdrawn, the question does not disappear, it sharpens. What remains, after the mourning, is the possibility of forging something: not a return to what came before the parenthesis, but a way of working that has passed through the parenthesis and come out changed. Somiya, the father in Late Spring, tells his daughter that happiness is not a state but a practice: “something you create yourself,” through effort, over years, with no guarantee that the effort will succeed. He is not describing a career plan. He is describing what it means to build a shared life out of nothing but your own willingness to keep building it, knowing that what you build will eventually be lost too.

The parenthesis of the job was shade. It was real, and it was temporary. What comes after it will require us to become, again, the kind of people who can answer the question “what do you do?” without pointing to a building. You mourn the structure that was holding you, not because it was perfect, but because it was yours, and it is going. The blacksmith knew what he made. The question the closing parenthesis puts to the rest of us is not whether we can go back to his world. We cannot. The question is whether we can do what he did: name the work, and mean it. Thoreau’s Walden ends with the assertion: “The sun is but a morning star.” What we took for the full light was only the beginning. The parenthesis was shade. The morning after it will be harder, brighter, and ours.

---

Sources & further reading

Ronald Coase, “The Nature of the Firm” (1937). The foundational economic argument that firms and employment exist primarily to minimize the transaction costs of coordinating labor through the open market.

E.P. Thompson, “Time, Work-Discipline, and Industrial Capitalism” (Past & Present, 1967). A historical account of how the rise of industrial manufacturing transformed the human experience of work from task-orientation to clock-time discipline.

Hannah Arendt, The Human Condition (1958). A philosophical framework distinguishing labor, work, and action to examine what is made, what is maintained, and how human identity is disclosed.

Richard Sennett, The Craftsman (2008). An argument that making is thinking and that intelligence is embedded within the act of physical creation, challenging Arendt’s sharp division between work and contemplation.

Stanley Cavell, The Senses of Walden (1972/1992). A reading of Thoreau’s Walden that exposes the acoustic identity of morning and mourning as the structural key to awakening and self-relinquishment.

Stanley Cavell, In Quest of the Ordinary (1988). An exploration of the ordinary as a lost condition that we must continuously mourn and maintain rather than assume as a permanent possession.

David Graeber, Bullshit Jobs (2018). A sociological analysis of modern corporate jobs that have lost their connection to meaningful production, highlighting the administrative drowse inside the firm.

Hubert Dreyfus, Being-in-the-World (1991). A commentary on Heidegger’s Being and Time arguing that skill acquisition is embodied and situated, requiring stages of struggle that cannot be automated.

Jônadas Techio, “Our Late Spring” (2026). A companion essay on film and loss that develops the concept of mourning as a creative practice of letting go and beginning again.

Jônadas Techio, “What Won’t Cross” (2026). An essay exploring the limits of transmissible knowledge, arguing that genuine thinking cannot be outsourced to tools but must be lived as a form of formation.

Jônadas Techio, “Calibration Debt” (2026). An essay on the hidden cost of delegating judgment, examining how organizations accumulate cognitive debt when systems absorb the friction of decision-making.

Domestic infiltrations

A room in postwar Japan. Tatami mats, kimonos, the mandatory seiza posture, a brazier heating water for tea. Women arrange themselves around the ceremony with studied care. Every gesture belongs to a tradition older than anyone present.

Then a young woman enters carrying a modern purse.

It is a small thing, Western-style, and she becomes self-conscious about it almost immediately. This object does not belong here. She tucks it beside her, half-hidden. Nobody else seems to notice. The ceremony continues. But the incongruity has been registered, and Yasujiro Ozu’s camera has made sure we registered it too.

The young woman is Noriko, played by Setsuko Hara with a radiance that makes her eventual sadness almost unbearable. The film was made during the American occupation of Japan, two years after a new Constitution had come into force and reshaped Japanese political and social life. It is saturated with small incongruities like the purse: a “Drink Coca-Cola” sign glimpsed during a bicycle ride, a “Time-Life” billboard in Tokyo, a young scholar wearing a Western suit while his employer sits in a kimono. None of these are catastrophic. The new world does not announce itself. It simply appears, alongside the old one, and waits.

Ozu knew nothing about language models. But he understood what it looks like when a form of life is quietly replaced by another. And he understood something else, something I think the current discourse about AI has not yet learned: that the proper response to such a replacement is neither panic nor acceptance. It is something more difficult than either. The response is mourning, but in a sense that has almost nothing to do with grief.

The weight of tea

At the center of the film is a relationship between a widowed professor, Shukichi Somiya, and his twenty-seven-year-old daughter Noriko, who keeps his household running. Ozu gives the first half of the film to this relationship, with a patience that can feel, to a viewer trained on Western pacing, like nothing is happening. We watch Noriko prepare dinner. We watch her make tea. We watch her tease her father about forgetting things, hang up his clothes, move through the rooms of their house with a freedom that is also a form of devotion. He depends on her entirely, and the dependence runs both ways.

The world they have built is not grand. It is made of tea and tidying and familiar jokes. It absorbs their attention, their care, their identities. It is a world. And Ozu spends an entire hour making us feel its weight so that we will feel its loss.

In one of the film’s early scenes, Somiya sits on tatami in his traditional house, working with his assistant Hattori on a paper about the German-American economist Friedrich List. Hattori wears a Western suit. The research belongs to a Western academic tradition. The house is Japanese.

Two forms of knowledge share the same room, and neither has yet displaced the other. But the displacement is coming, and the professor’s body, positioned between tradition and modernity, already registers it. He is maintaining a community of judgment through embodied, situated, intellectually alive work: a man in a kimono, on a tatami floor, doing Western economics. The incongruity is Ozu’s signature.

Stanley Cavell once wrote that “any relationship of absorbing importance will form a world.” What holds this one together, what holds any shared world together, is not a system or a structure. It is the daily, embodied attention of the people inside it. The willingness to notice that a sentence is not quite right and to sit with the discomfort until you understand why. The patience to read a student’s work slowly enough to feel where the thinking breaks down. The habit of care that makes someone check a source one more time, not because a rule demands it but because their name is attached to the result. Noriko makes tea. The world holds.

Persistence without force

The film’s turning point arrives during a Noh performance. Somiya and Noriko attend together. Noh is a practice of extreme formalization: precise movements expressing deep emotion through minimal gesture. There is an apt parallel with Ozu’s own filmmaking, which conveys its most profound feelings through the smallest shifts in posture or expression. The audience, in kimonos and suits, sits in respectful attention.

During the performance, Noriko notices a woman she knows: Mrs. Miwa, the widow who is being considered as a new wife for her father. Noriko’s response is barely visible. A subtle lowering of the head. A withdrawal into herself. It is enough.

In that gesture, the daughter registers her father’s separateness for the first time: the fact that his desires may not be aligned with hers, that his happiness might require her loss, that the world they share is not guaranteed. She is discovering that the person closest to her is also, irreducibly, another person, with needs she cannot absorb into her own.

What stays with me now is what this scene also says about ritual itself. The Noh play continues. The forms are intact. But something has shifted in how the participants inhabit them. I once wrote that the traditional practices Ozu depicts are at the verge of becoming mere ritualized repetitions, or perhaps forms of mourning a world already felt as lost. The ritual does not vanish. It continues, hollowed, while the conditions that gave it meaning shift beneath it. The Noh play is still beautiful. Noriko is still sitting in the audience. But the world the play was made for, the world in which a father and daughter could sit together in untroubled intimacy and watch performers enact the old stories, is already receding.

Anyone who teaches, or has recently been a student, will recognize the feeling. A lecture hall in which the professor speaks and the students take notes still looks the same as it did twenty years ago. The room is the same. The postures are the same. The ritual persists. The world it was made for is receding, and, as I argued in What Won’t Cross, the slow, formative practice of thinking does not automatically cross to a new medium. The students are half-present, their attention divided between the lecture and the devices in their laps, and the professor senses this without always being able to name it, and adjusts, simplifies, performs a version of teaching that accommodates the drift.

This is what persistence without force looks like. And this is where the film first teaches us what mourning might mean. Not grief for someone who has died, but the recognition that the world you inhabit with another person is not permanent. Love does not make two people one. Separateness is the condition of every genuine relationship. Noriko’s lowered head is the gesture of someone who is beginning to accept this, and the acceptance will cost her everything she has, and it is also, Ozu insists, the only path toward an adult life. Mourning is not the opposite of growth. It is its prerequisite.

Beyond ceremony

Late Spring builds toward a wedding that never appears on screen.

We see the preparation. Noriko stands in her room, now emptied of furniture, dressed in a traditional wedding kimono. Her face holds an expression that Setsuko Hara somehow makes legible as contentment, sadness, gratitude, and farewell simultaneously. She kneels before her father and thanks him for his care. He smiles and tells her to be happy. It is the last time we see them together.

Then Ozu cuts away. The wedding ceremony happens offscreen. We are not shown it. We will never be shown it.

This is the film’s most radical gesture. A wedding is where shared meaning does its most visible, most ancient work. Two people stand before a witness. A community recognizes the act. A private relationship becomes a public fact. Every culture on earth has some version of this moment, and every version rests on the same fragile structure: it works because the people present agree to treat it as real, because they invest the form with their collective belief, because someone stands behind the words and means them. Take away the standing-behind and the ceremony becomes mere performance. The words are still spoken. The rice is still thrown. But the thing that made it a wedding (the human willingness to be bound) has quietly departed.

Ozu refuses to show the ceremony. Cavell helps me understand why. In his study of Hollywood remarriage comedies, Cavell noticed that classical comedy always ended with a wedding, celebrating society’s ability to provide continuity. But at some point, he wrote, “perhaps when the world went to war, society stopped believing in its ability to provide that continuity.” The remarriage comedies respond by locating happiness not in ceremony but in the couple’s own “capacities for improvising a world, beyond ceremony.” The guarantee is gone. What remains is the difficult, daily, never-finished work of forging a shared life out of nothing but your own willingness to keep forging it.

On the night before the wedding, in Kyoto, Noriko tells her father that she does not want anything to change. “Being with you is enough for me,” she says. “Marriage couldn’t make me happier.” Somiya’s reply is one of the most remarkable speeches in all of cinema:

“Happiness isn’t something you wait around for. It’s something you create yourself. Getting married isn’t happiness. Happiness lies in the forging of a new life shared together. It may take a year or two, maybe even five or ten. Happiness comes only through effort. Only then can you claim to be man and wife.”

A father who has spent the entire film being cared for by his daughter now tells her the deepest thing he knows: that happiness is not a state but a practice, not a destination but a labor, and that the labor takes years, and the years will be difficult, and the difficulty is the point. He is telling her what mourning actually opens onto. You accept that the world you had is gone. You accept that no ceremony can replace it. And then you begin the slow, uncertain work of forging a new one, knowing that this new world too will eventually be lost, and that the willingness to forge it anyway is the only happiness available to finite beings.

This is the connection between Ozu’s film and our present that I find hardest to articulate, and most important to try. Somiya is not describing efficiency or optimization: the very forces that, when execution goes to zero, expand our orchestration to infinity (as I examined in The Jevons Paradox of the Self). He is describing what it takes to maintain a shared world: effort, over time, by beings who cannot know in advance whether the effort will succeed.

When AI systems take over the maintenance of institutional reality, that effort disappears. The outputs continue. The forms remain intact. Legal filings appear, correctly formatted. Medical summaries read as they should. Policy documents, research papers, editorials, all produced with impressive competence. But the system that produced the document does not bear the risk of its failure. It does not have a reputation that can be destroyed, a patient whose face it will have to meet, a professional community whose quiet judgment it fears. It does not need a year, or two, or ten. It does not forge anything. What we get is the ceremony without anyone standing behind it. The form without the forging. The wedding, performed to an empty room.

What the peel knows

After the wedding, Somiya returns home alone.

Ozu films him entering the house, standing for a moment in the doorway as if the space has become unfamiliar. He sits in a chair that used to be his daughter’s. The camera lingers on the empty rooms, the rearranged furniture, the corridors where Noriko used to move with such freedom. The absence is not stated. It is architectural. The house is the same house. It is no longer the same world.

Then comes what I believe is one of the most extraordinary uses of an ordinary scene in the history of cinema. Somiya sits and peels an apple. Carefully, slowly, trying to keep the peel in one piece. He fails. The peel breaks. He lowers his head.

The film does not cut to the peel on the floor. The critic Andrew Klevan observed that the film does not cut to the peel falling onto the floor because it wishes to maintain the integrity of something lost from the frame and now unrecoverable. The camera stays on the man, on the failure, on the lowered head. It is the same movement Noriko made at the Noh play when she registered her father’s separateness. Father and daughter, in their separate moments of loss, make the same gesture. The echo is devastating.

The image is itself the argument: Somiya tries to keep the apple peel whole and fails. The peel falls and the film refuses to show us where it lands. What is being depicted is not an event but a condition: the condition of beings who try to hold together what is already falling apart, and who must go on living after it falls. This is mourning in its full sense: not grief, not nostalgia for a golden age, not a clinical stage that resolves into acceptance, but something more precise and more demanding.

Cavell, drawing on Thoreau, describes mourning as the path of accepting the loss of the world, accepting it as something which exists for us only in its loss. The everyday is not a stable ground beneath our feet. It is something we must continuously achieve, continuously maintain, and continuously mourn, because it is always already receding from us. He condensed it into a sentence I have returned to more times than I can count: “The everyday is what we cannot but aspire to, since it appears to us as lost to us.”

This is not the familiar mourning that resolves when you find a replacement for what you have lost. It is an ongoing reckoning with the fact that the world is separate from us, that our connection to it is not guaranteed, and that maintaining that connection is our responsibility, not the world’s gift.

The happiness speech and the apple peel scene make the same thing visible from different angles. Mourning and forging are not opposites. They are two faces of a single activity. To forge a new shared world, as Somiya tells Noriko she must, you have to accept the loss of the old one. To accept the loss, truly, you have to be the kind of being for whom the loss is real, for whom the world was not a backdrop but a home, for whom the peel breaking is not a mechanical event but a private catastrophe. Mourning is what makes forging possible. It is the cost of admission to a genuinely shared life.

Somiya, peeling his apple alone in a house that is no longer his daughter’s home, inhabits this condition. He is not trying to get Noriko back. He sent her away. He is living inside the loss and learning, at fifty-six, what the ordinary costs when it is genuinely borne.

This is what AI cannot do. Not because it lacks feelings (that question remains open in ways that should make us uncomfortable). But because mourning requires finitude. It requires the capacity to lose something irreplaceable and to know it is irreplaceable. It requires a body that sits in a chair that used to be someone else’s, hands that try to keep an apple peel whole and fail, a head that lowers when the failure becomes real. A system that maintains shared reality without having a stake in its coherence, that keeps the forms running without knowing what it would mean for the forms to collapse: such a system cannot mourn. And because it cannot mourn, it cannot forge. It can produce, optimize, maintain, generate. But the forging that Somiya describes to his daughter, the labor of years, the happiness that comes only through effort: this requires a being that has something to lose. The shared infrastructure of meaning that holds civilization together was built by such beings. It was built by mourners.

I build these systems. I know their power.

And I know the difference between maintaining something and caring about it is the difference between a civilization that knows what it has and one that merely has it.

The waves at Kamakura

Late Spring ends with a shot of the ocean. Waves break on the beach at Kamakura. They are indifferent to everything that has happened. They broke before Somiya was born. They will break after everyone who watched the film is gone.

The deepest risk AI poses to our shared world is not catastrophe. But it is also not inevitable decline. This is the part that Ozu’s film understands and that most of the current discourse does not.

Noriko leaves. Somiya is alone. The apple peel breaks. And then: the waves. The film does not end with the father in his empty house. It ends with the ocean. With something that goes on.

Mourning is not a posture of defeat. Somiya told his daughter the truth: happiness comes only through effort, through the forging of a new shared life. He was preparing her not for loss but for what comes after it. For the difficult, uncertain, necessary work of building a world she cannot yet imagine. The mourning makes the forging possible. Without it, she would cling to the old arrangement forever. And the old arrangement, however beautiful, was a world in which she never had to become herself.

This is what the AI discourse is missing. Not just the recognition that something is being lost, though that recognition is urgent, and mostly absent. The recognition that the loss, properly mourned, could open onto something. The point is not that human error is valuable in itself, or that inefficiency deserves protection. It is that institutional reality is written into human flesh through the experience of getting it wrong.

A doctor who misdiagnoses a patient carries that error in her body. It is not abstract regret; it is a specific face returning at three in the morning when sleep will not come. It changes how she reads the next scan: slower, with a peripheral attention that no protocol taught her. It changes how she listens, leaning forward when a patient hesitates, hearing the thing that almost went unsaid. That weight is not a flaw in the system; it is the system. A lawyer who misses a precedent, a teacher who grades too generously: each carries the cost forward, and the cost becomes part of their attention, part of what makes institutional reality real. These are not inefficiencies. They are the scar tissue of forging. They are what mourning makes available.

A system that does this work without bearing these costs can be more accurate than any human practitioner and still hollow the institution it serves.

I know, because I helped design the purse.

The excellence is the problem: it makes the costs look like overhead, makes the standing-behind look like a bottleneck, makes the human weight of institutional maintenance look like something that will be engineered away.

And yet. The doctor who has worked alongside an AI diagnostic system and who understands, viscerally, what it means to carry a patient’s face home: that doctor practices medicine differently. Not worse, not the same, but with a kind of attention that has passed through the fire of comparison and come out changed. The teacher who has felt the ritual of the classroom hollow itself out and who has decided, nevertheless, to keep teaching, to find new ways to make the encounter real: that teacher is forging something. The lawyer, the journalist, the researcher who has reckoned honestly with what these systems can do and has asked, therefore, what only a finite, fallible, mortal practitioner can bring to the work: these people are not relics. They are mourners. And mourners, Ozu shows us, are the ones who build what comes next.

We are in our own late spring. The old forms of maintaining shared reality are still here. The new ones have arrived alongside them, quietly, competently. The purse is already at the tea ceremony. The question is not whether the transition will happen. The question is whether we will mourn it well. Whether we will do what Noriko does, what Somiya does, what the film itself does: accept the loss without pretending it is not a loss, and then begin the slow, uncertain, necessary work of forging the world that comes after. The waves break at Kamakura. They do not grieve. They do not forge. They simply continue. We are not waves. That is our burden. Ozu knew it is also our grace.

---

Sources & further reading

Yasujiro Ozu, Late Spring (Banshun, 1949). A film about what it costs to let go of the ordinary, told through the quiet dissolution of a father-daughter household in postwar Japan.

Stanley Cavell, In Quest of the Ordinary: Lines of Skepticism and Romanticism (1988). Contains the formulation “the everyday is what we cannot but aspire to, since it appears to us as lost to us” (p. 171), and the account of mourning as the path of accepting the world in its loss (p. 172).

Stanley Cavell, Pursuits of Happiness: The Hollywood Comedy of Remarriage (1981). The argument that happiness must be found through the couple’s “capacities for improvising a world, beyond ceremony” (p. 239).

Stanley Cavell, Must We Mean What We Say? (1976). Contains the formulation “any relationship of absorbing importance will form a world” (p. 118).

Jônadas Techio, “Mourning the Loss of the Ordinary: A Cavellian Reading of Ozu’s Late Spring“ (Aesthetic Investigations, 2020). The scholarly foundation for the film reading developed here, focusing on Cavell’s diagnosis of skepticism and its cinematic expression.

John Searle, The Construction of Social Reality (1995). The formalization of institutional facts and the “X counts as Y in context C” structure that illuminates the absent wedding scene.

Andrew Klevan, Disclosure of the Everyday: Undramatic Achievement in Narrative Film (2000). The finest close reading of the film’s visual grammar, including the observation that the film “wishes to maintain the integrity of something lost from the frame and now unrecoverable” (p. 159).

Tammy Clewell, “Cavell and the Endless Mourning of Skepticism” (Angelaki, 2004). The distinction between Freudian mourning (substitution) and Thoreauvian mourning (accepting the loss without recovery), which grounds the argument about AI and finitude.

Nobody reassigned her. That is the part I keep returning to.

A marketing manager at the company where I work built a plugin that applies the company’s voice to every output every AI agent produces, across every function, automatically. She did not build a tool that helps other people write better copy. She encoded the brand’s vocabulary rules, tone constraints, and formatting conventions into an instruction file, and connected it to the shared agent layer. After that, a finance analyst drafting a summary, an investigator formatting a case report, an executive preparing a board update: the output arrives already in voice. Nobody invokes the plugin. Nobody routes a draft through marketing for review. The work that used to require marketing’s hands now happens at the infrastructure level, silently, everywhere.

She was not reassigned because nothing, on paper, changed. Her title is the same. Her reporting line is the same. The org chart still shows marketing as the function responsible for brand expression. But brand expression is no longer a task that flows through her team. It is a property of the system her team built. The chart says marketing does the brand. The truth is that marketing governs the brand, and the execution has dissolved into something no department owns.

There is a distinction we learn early in school and forget by the time we enter an office. Can describes ability. May describes permission. “Can I leave the room?” draws a correction: “You can leave. The question is whether you may.”

For most of the twentieth century, corporate life made the distinction invisible. A customer success manager could not build a branded commercial proposal because building one required access to design tools, knowledge of the brand guidelines, familiarity with pricing structures, and hours of formatting work. The permission boundary and the competence boundary overlapped almost perfectly. The org chart was not enforcing an arbitrary rule. It was describing a real limit: what a person in one function could accomplish without the tools, training, and context held by another.

When the cost of execution drops toward zero, the competence boundary dissolves. The permission boundary remains. The chart becomes a skeuomorph: the leather stitching on the digital calendar, the wood-grain shelf in the e-reader app. A visual memory of a constraint that no longer exists, maintained because the organization has not yet noticed that the constraint was the only thing holding the metaphor together.

What permission used to cost

John Searle spent decades studying the strange objects that hold civilization together: money, property, marriage, borders, corporations. He called them institutional facts. An institutional fact exists when a community agrees, usually without anyone having explicitly agreed, that X counts as Y in a given context. A piece of paper counts as legal tender. A signature counts as consent. A title on a business card counts as authority to approve a purchase order.

Institutional facts are powerful because they are invisible when they work. You do not pause, each time you buy a coffee, to marvel at the collective hallucination that makes a piece of polymer worth four dollars. The fiction is seamless. It operates below the threshold of attention.

The org chart is an institutional fact of this kind. “Marketing” counts as the function responsible for brand voice. “Pre-sales” counts as the function responsible for proposals. “Intelligence” counts as the function responsible for threat analysis. These assignments were never descriptions of competence. They were descriptions of permission. And the permission was credible because the competence required to cross a functional boundary was enormous. You could not do marketing’s job without marketing’s knowledge. The chart was load-bearing because the walls it drew corresponded, roughly, to real limits on what anyone outside the function could accomplish.

When the walls lose their load, the fact that depends on them begins to float. It is still on the chart. It is still enforced by reporting lines and budget codes. But it has detached from the reality it was supposed to describe, the way a price tag stays on a shirt after the store has closed. The tag is still there. The store is not.

The function that dissolved itself

At Axur (now Infoblox), the cybersecurity company where I work, twenty-three people across eight functions spent seven weeks building a shared knowledge architecture on top of AI agents and a Git repository. None of them were engineers. I helped design the architecture and co-wrote the manifesto that framed it with the company’s CEO. What I want to describe here is not the system (you can read about it in the link above). It is what the system revealed about the chart.

The marketing manager’s plugin is the clearest case (just one of many!). She did not automate her team’s workflow. She made her team’s gatekeeping function unnecessary. Before the plugin, every piece of external communication passed through marketing for brand review: the tone was wrong, the vocabulary was off, the formatting did not match. Marketing was a bottleneck, and the bottleneck was the institutional fact that made the function legible. “Marketing reviews brand expression” was how the chart justified marketing’s place in the workflow. The review was the load the wall carried.

The plugin dissolved the bottleneck by encoding the review into the infrastructure. The brand rules are now applied automatically, upstream, before anyone sees the output. Marketing no longer reviews brand expression for the same reason you no longer check a dictionary to spell-check a document: the correction happens at a layer below conscious attention. The function’s most visible task, the one the chart used to point at as proof the function was needed, evaporated.

What remained was something harder to see and more valuable. Marketing still owns the definition of the brand: the vocabulary, the tone, the constraints. Marketing updates the plugin when the voice evolves. Marketing decides what the rules are. The function did not disappear. It shifted from performing the brand to governing it. And governance, unlike performance, does not require a bottleneck. It requires authority, taste, and a file that every agent reads.

The same pattern appeared across the pilot. A customer success contributor built a skill that generates branded proposals by cross-referencing internal product knowledge, a task that had belonged to pre-sales. A product manager encoded an entire analytical methodology into six instruction files, a task that had belonged to intelligence. An operations team built investigation dossiers that crossed into intelligence’s territory. In every case, nobody changed the chart. The competence boundary dissolved. The permission boundary stood. The gap between the two widened until the chart became decoration.

From execution to governance

The temptation, when you see departmental walls dissolving, is to reach for the familiar narrative: hierarchy is dead, the future is flat, everyone does everything. That story has been told before. It has never worked. HubSpot experimented with removing titles and org charts and reverted because the coordination cost without structure was unsustainable. Holacracy promised self-organization and produced, in most implementations, a different and more opaque kind of bureaucracy. Every attempt to flatten the hierarchy by removing it has collided with the same problem: people need to know who decides, who reviews, who is accountable when something goes wrong.

The skeuomorphic organization does not need to be flattened. It needs to be made honest.

The chart can keep its reporting lines, its accountability structures, its decision rights. What it cannot keep is the pretense that functional boundaries correspond to competence boundaries. The shift is from ownership of execution to ownership of methodology. The department that understands how something should be done encodes that understanding into the shared layer. Every other function composes against it. The chart does not disappear. Its edges become permeable. What crosses the boundary is not people, not reporting lines, not authority. What crosses is encoded competence: methodology made legible to agents and to every function that reads the same files.

The institutional fact “marketing owns brand expression” was true when brand expression required marketing’s hands. When brand expression becomes a property of the infrastructure, the fact does not become false. It becomes a different kind of fact: a fact about governance rather than about labor. The distinction sounds administrative. It is structural, and most organizations have not made it yet.

This shift, once you see it, explains something that puzzled me during the pilot. The people who moved fastest from using AI to building infrastructure were not engineers. They were the people closest to the methodology: the analyst who understood the investigative workflow, the product manager who had spent years learning the threat taxonomy, the marketing manager who knew exactly what the brand should sound like. They encoded what they knew because the encoding surface existed (a shared repository, a thin manifest, plain-text instruction files) and the knowledge was already in their heads. Engineering was not required. Understanding was. The people closest to the work were the ones best positioned to encode it, and any company that waits for engineering to lead this transition will discover that it waited for the wrong function.

But the permeability also creates a trust problem. When anyone in any function can compose against the encoded methodology of any other function, the question of trust becomes urgent. Who changed the product specification? When? Can we see the version before the change? Can we roll it back?

The pilot converged, without anyone planning it, on Git: version-controlled plain text, not a database, not a SaaS platform. The convergence was not an engineering preference. It was a governance discovery. The skeuomorphic organization could afford opaque knowledge flows because the walls ensured that only the authorized function touched the authorized artifacts. The permeable organization cannot afford opacity. Trust is downstream of traceability. Traceability is what version control is for.

The weight that was not transferred

I should name the part of this that keeps me awake.

In What Won’t Cross, I argued that the formation stage of intellectual work does not transfer to the machine. The artifact crosses; the person who was being shaped by making it does not. A student who lets the AI write the problem set gets the grade and loses the learning. The product is identical. The formation is absent.

The same argument applies at the institutional level, and it is more dangerous there because institutions are less likely to notice.

When the marketing manager encoded the brand voice into a plugin, something valuable happened: every function in the company gained access to methodology that had previously lived in three people’s heads. Something else happened, more quietly: the next person who joins marketing will never learn the brand voice through the slow, tedious process of reviewing other people’s drafts and correcting them by hand. The formation stage, the months of reading bad copy and feeling the wrongness before articulating it, the pattern recognition that builds only through repeated exposure to off-brand language, all of that is now encoded in an instruction file that runs in milliseconds. The new hire will govern the plugin. She will not have been formed by the practice the plugin replaced.

You have seen this if you manage a team. A junior employee joins and immediately produces polished, competent work because the AI handles the execution. The output is indistinguishable from what your best people produce. But you watch the new hire in a meeting where something unexpected happens, where the client pushes back in a way the template did not anticipate, where the data contradicts the methodology, and you see the gap. The junior does not know what to do when the encoded path runs out. She has operated the pipeline. She has not been formed by the practice that built it. The competence is borrowed. It lives in the infrastructure, not in her.

Hubert Dreyfus spent his career arguing that expertise is formed through embodied, situated practice: the novice follows rules; the expert has internalized the practice so deeply that the rules become invisible. The expert acts from a kind of perceptual recognition that cannot be captured in a manual. The agentic architecture is optimized to eliminate exactly the inefficiency that Dreyfus’s model depends on. The encoded methodology skips the novice from rule-following to competent output without passing through the stages where expertise actually forms. The pipeline is better. The formation pipeline is broken.

I do not have a solution. The standard response is “the human learns higher-order skills instead,” but this assumes that higher-order skills can be formed without the practice of lower-order work. My experience supervising students for a decade tells me this is wrong more often than the optimists want to admit. You do not learn to make judgment calls by watching someone else’s judgment calls execute as automated workflows. You learn by making bad calls, sitting with the consequences, and carrying the weight of having been wrong in a way that reshapes how you approach the next case.

The skeuomorphic organization is being replaced by something new: a permeable firm in which encoded methodology flows across boundaries that used to be walls. The question that should follow every celebration of this fact is: where will the next generation of methodology-writers come from, if the practice that formed them has been encoded out of the path?

The ceremony and the weight

A wedding in which both people mean the vows is an institutional fact held together by human commitment. A wedding in which the ceremony is performed perfectly but nobody means anything by it is a skeuomorph: the form of the old thing, emptied of the weight that made it real.

The org chart, when it described real constraints, was a living institution. It will not disappear. Reporting lines serve purposes the permeable firm does not replace: accountability, escalation, career development, the human need to know who to talk to when things go wrong. The chart will persist as a social structure long after it has ceased to describe how work moves. The question is whether the generation formed by the old constraints can encode their judgment fast enough for the generation that will never encounter the constraints at all.

I do not know the answer. I helped build the system that is asking the question.

*

The resources from the Axur pilot are open-sourced on GitHub. You can read the full manifesto, browse the companion field notes, or clone the starter kit to experiment with the architecture yourself.

---

Sources & further reading

John Searle, The Construction of Social Reality (1995). The foundational account of institutional facts: the objects (money, marriage, property, borders, corporations) that exist because a community agrees they do, and that dissolve when the agreement withdraws.

Hubert Dreyfus, Being-in-the-World (1991) and “A Phenomenology of Skill Acquisition as the Basis for a Merleau-Pontian Non-representationalist Cognitive Science” (2002). The five-stage model of skill acquisition and the argument that expertise is embodied, situated, and cannot be captured in explicit rules. The formation-pipeline problem in this essay draws on Dreyfus’s account of what is lost when novices skip the stages that produce experts.

Fábio Ramos and Jônadas Techio, Restructuring Knowledge Work for Human-Agent Teams (2026). The manifesto and companion field report from the pilot described in this essay. Six principles, seven conventions, twenty-three contributors, eight functions, seven weeks.

Block, “From Hierarchy to Intelligence” (2026). Block’s description of a company-scale architecture adjacent to the one described here, including the “world model” concept of continuous organizational memory.

Jônadas, The Machine-Readable Self (2026). Companion essay on operational legibility: how work became machine-readable before AI arrived, and what the clearing reveals when the operational is finally delegated.

Jônadas, What Won’t Cross (2026). The formation-stage argument: the artifact crosses to the new medium; the person who was being formed by making it does not.

Jônadas, What I Built When Chat Stopped Being Enough (2026). The personal-workspace counterpart to the organizational architecture described here: the Agentic Studio, progressive disclosure, and the style-guide-as-software.

I wanted to build a knowledge base for the product at my company. A structured, version-controlled repository of everything the product knows, organized into Markdown files and tracked in GitHub. A year ago this would have been a quarter-long project: documentation sprints, alignment meetings, manual synthesis across teams and data sources. Instead I sat down with an AI and started distilling. I fed it product documentation, internal wikis, support tickets. It extracted structural claims, organized them into interlinked files, and I reviewed, corrected, and governed the output. Within days I had a working knowledge architecture.

I had been building a similar system for my own intellectual work, a governed note-linking system compiled from twenty years of reading notes, lecture fragments, and philosophical drafts. The two projects started reinforcing each other. The same architectural instincts that structured the product knowledge base sharpened the personal one. The philosophical reading I was doing in the evenings kept generating insights I could bring to work in the morning. I started using the AI as a sparring partner for my essays, arguing about Heidegger and Girard after dinner, then waking up to apply the same thinking to product intelligence. I wrote personal web applications and writing tools to test whether I could. Once the loop started, I could not stop. Each completed project revealed three more that were now plausible. Each solved problem surfaced a new frontier. At no point did I arrive at a natural stopping place.

The cultural narrative tells us that artificial intelligence is a labor-saving technology. The machine that will finally deliver the post-work utopia, or at least Keynes’ long-promised fifteen-hour work week. The assumption underneath is that human desire is a fixed bucket. If it takes forty hours to fill the bucket today, and the machine can fill it in four, we will take thirty-six hours of leisure.

My experience suggests the opposite. The bucket has no bottom. And we are not about to work less.

Jevons’s coal

In 1865, the economist William Stanley Jevons published The Coal Question, in which he noticed something counterintuitive about the steam engine. James Watt’s improvements had made steam engines vastly more efficient. The reasonable expectation was that coal consumption would decline: better engines, less fuel per unit of work. What Jevons observed was the opposite. Total coal consumption in Britain was skyrocketing. Efficiency had not satisfied the demand for coal. It had unlocked demand that nobody knew existed. Steam power, suddenly cheap enough to be profitable in a thousand new applications, was being applied to industries that had never used it before. The more efficient the engine became, the more coal the economy burned.

This pattern, now called the Jevons Paradox, keeps showing up whenever a resource becomes dramatically cheaper to use. We do not conserve it. We find new, previously unthinkable uses for it, and total consumption increases. When data storage became virtually free, we did not stop buying hard drives; we started recording our entire lives in high definition. When spreadsheet software made financial calculations instantaneous, it did not eliminate the need for accountants; it spawned an entire industry of quantitative modeling that demanded vastly more calculation.

What I have been living through, sitting at my desk after hours building intelligence architectures that did not exist a month ago, is the Jevons Paradox applied to human ambition.

When the cost of cognitive execution drops toward zero, we do not accomplish our usual tasks and go home. We attempt things that were previously unthinkable. The machine does not replace our effort. It makes our effort so wildly productive that we cannot stop. The lever moves so easily now that the only question left is whether your arms give out.

The bottleneck shifts

In a conversation with Dwarkesh Patel in January 2025, the economist Tyler Cowen described this structural shift with characteristic directness: when AI removes the technical constraints on execution, we become the bottleneck. He was talking about institutions, about how regulatory bodies and cultural inertia would throttle the speed of AI’s impact. I have been writing about Cowen’s broader project elsewhere, in The Claim Upon the Training Data, where I examined his practice of writing for AI training indexes, and in What Won’t Cross, where I used his argument about unassisted writing as the foundation for a larger claim about cognitive formation. But the bottleneck observation is the one that has followed me home.

I have been feeling it in my body. When the tools can build, write, and synthesize at the speed of thought, the only limit remaining is how fast I can think. The bottleneck is my need to sleep. My capacity to hold the architecture in my head. My stamina for sustained orchestration across a dozen simultaneous projects. The technical constraint was a ceiling. The ceiling is gone. What remains is the floor: the biological minimum below which I stop functioning.

Both sides of the AI labor debate share the same assumption and get the same thing wrong. The utopians promising leisure and the doomsayers predicting mass unemployment both treat the demand for human effort as a fixed quantity. The machine either fills it, freeing us, or replaces it, destroying us. Neither side considers the possibility that the machine might expand it. That when execution becomes cheap, the appetite for what can be executed does not stay constant.

It explodes.

The other response

I have described one side of what happens when execution becomes frictionless: the person already biased toward building discovers that the lever now moves without resistance, and pulls it until something in the body gives. That has been my experience, and I suspect it is the experience of many readers of this essay.

There is another response, and it is already visible at scale.

In 2025, writing projects on major freelance platforms declined by roughly a third year over year. A survey of nearly seven thousand creative professionals found that a third of illustrators had lost commissions directly to generative AI. Translators were among the hardest hit. The statistics describe job loss, but the reports underneath describe a psychological toll deeper than unemployment. They describe demoralization. Artists who still have work but have stopped wanting to make it. Writers who can still write but feel something draining out of the act. The word that keeps appearing in these accounts is not “replaced.” It is “pointless.”

This is worth pausing on, because the feeling of pointlessness is not a rational calculation about labor markets. A writer whose income drops can retrain, pivot, adapt. The demoralization runs deeper. It is the experience of watching a machine produce, in seconds, a version of the work that took you years to learn. The output may be worse. It may lack the texture that makes your work yours. It does not matter. The sheer effortlessness of the machine’s production changes the way your own effort feels. What used to feel like craft now feels like friction. What used to feel like earned skill now feels like an inefficiency the market is about to correct.

The philosopher René Girard spent his career studying this structure. In Deceit, Desire, and the Novel (1965), he argued that human desire is fundamentally mimetic: we do not generate our wants from within. We absorb them from others. We want what we see others wanting, and we build what we see others building. In The Economics of Infinite Desire, I explored Girard’s framework at length and traced its consequences through Silicon Valley, Versailles, and the structure of academic life. The key insight for this essay is simpler: desire requires a model. We need to see someone doing the work, struggling with it, making it look possible and worth pursuing. The model shows us not just the desirability of the goal but the plausibility of reaching it as a human activity.

What happens when the model is a machine that does not struggle?

For builders, the machine acts as what I called a “spotter” in The Friction That Was Thinking: a training partner in a cognitive gym, someone who stands behind the bench while you press, who does not lift the weight for you but lets you attempt heavier lifts than you could manage alone. The frictionless AI makes building harder to resist, not easier to abandon.

For others, the encounter with effortless machine execution sends the opposite signal. If desire is borrowed, and the model you were imitating has been replaced by something that makes human participation look vestigial, then the desire to build does not survive the comparison. Girard would call this a thinning of the desire: a want that collapses the moment it is seriously challenged. The writer does not lose the ability to write. The writer loses the sense that writing, as a human act, still means something against the backdrop of a machine that can produce infinite text without effort or intent.

We may be witnessing the early stage of a bifurcation that will define the next decade: between those who use the machine to scale their orchestration until their biology fails, and those who look at the frictionless output and quietly stop wanting to make things at all.

Standing-reserve

There is something I have not confessed yet about the euphoria.

Somewhere in the third week of this, I noticed that I had started resenting meals. Not the food. The interruption. Twenty minutes away from the screen felt like a system failure. I caught myself optimizing my sleep schedule, not for rest, but for output: calculating the minimum hours I could function on, treating my own alertness as a production input to be managed. My attention had become, in my own thinking, a resource to allocate.

I recognized this feeling. Not from personal experience. From philosophy.

Martin Heidegger, in his 1954 essay “The Question Concerning Technology,” described a way of relating to the world that he called Bestand, usually translated as “standing-reserve.” It is the orientation in which everything, including nature, labor, and human beings, becomes a resource held in readiness for further use. The coal is not a substance with its own existence. It is a fuel waiting to be burned. The river is not a waterway. It is a hydroelectric opportunity. Even the forester walking the same path his grandfather walked is, whether he knows it or not, subordinated to the orderability of cellulose for the paper industry.

Heidegger was describing a disposition, not a technology. A way of seeing the world in which nothing is allowed to simply be what it is. Everything is revealed as available, flexible, waiting to be put to work.

The euphoria of hyper-labor is the moment when this disposition turns inward. I was not resting in my finitude. I was mobilizing myself as a resource for my own projects. The tools had become so responsive, so frictionless, that the only remaining inefficiency in the system was my own body. And I was trying to optimize it away.

The joy was real. The productivity was real. The things I was building had genuine value. And underneath all of it was a logic that had no internal stopping point, no built-in signal that says enough, no mechanism by which the person pulling the lever learns to let go.

The discipline we do not have

In previous essays I have been mapping a set of related dangers. In What Won’t Cross, I argued that the formation stage of intellectual work, the hours where a person is being shaped by the act of making something, does not transfer to the machine: the artifact crosses to the new medium, the formation does not. In The Single-Player Game, I described how external scorekeeping can quietly replace the internal game you were actually playing, how the switch from one to the other does not announce itself. In The Economics of Simplified Living, I argued that the gesture of refusal, the “AI-free” badge, the performative rejection of the technology, is itself available for mimetic imitation and therefore is not a solution.

A different danger operates even when the work is genuine, even when the formation is intact, even when the score is internal. The danger is that there is no ceiling.

When the tools were slow, execution imposed a limit. You could not build more than the day allowed. The gap between what you wanted to do and what you could do was wide enough to enforce pauses. Those pauses were the spaces in which you remembered that you were a body, that you needed to eat and sleep and see people and do nothing for a while. The gap was never just an obstacle. It was a governor on the engine.

The governor is gone. The engine has not changed.

Every discipline humans have developed for managing excess was designed to resist temptation toward ease: gluttony, sloth, indulgence, distraction. Every spiritual tradition, every philosophy of moderation, assumes the danger is that we will do too little, consume too much, surrender to comfort. We have entire civilizational architectures for resisting the pull of laziness.

We have almost nothing for resisting the pull of productive intensity.

What does that discipline even look like? I do not think it looks like refusal. I do not think it looks like rationing screen time or setting timers or performing the aesthetics of digital minimalism. Those are responses to distraction, and the problem I am describing is the opposite of distraction. The problem is a focus so total that the body becomes invisible to the mind that inhabits it.

Jevons published The Coal Question in 1865. He named the paradox with precision: efficiency would not conserve the resource, it would accelerate its consumption. British coal production peaked in 1913 and continued at enormous scale for decades after. Nobody changed course. The paradox was identified, understood, discussed, and entirely ignored for nearly fifty years. Naming a trap has never, historically, been the same as escaping it.

I closed the laptop last night. Not because I had finished. I will not finish. The system has no endpoint, and the only thing that does is me. The coal does not know it is being burned. That is the one advantage I have, and I do not yet know whether it is enough.

---

Sources & further reading

William Stanley Jevons, The Coal Question (1865). The original observation that efficiency of resource use can increase total consumption. British coal production peaked in 1913, nearly fifty years after the warning.

Tyler Cowen, interview with Dwarkesh Patel (January 2025). The argument that humans and institutions become the bottleneck as AI removes technical constraints. Part of a broader body of work explored in The Claim Upon the Training Data and What Won’t Cross.

Eldar Maksymov, “The Jevons Paradox and Insatiable Humans: Why AI Won’t Empty the Finance Suite” (2026, SSRN preprint). An optimistic application of the Jevons Paradox to professional work, arguing that cheaper intelligence will expand demand for analysis. This essay takes the structural mechanism seriously while questioning whether expanded demand equates to human flourishing.

René Girard, Deceit, Desire, and the Novel (1965). The first systematic account of mimetic desire: we want what others want, and we mistake the imitation for autonomy. The bifurcation between hyper-labor and creative paralysis draws on Girard’s account of how desire thins when the model becomes unimitable. See also Luke Burgis, Wanting (2021), for the most accessible contemporary treatment. I explored Girard’s framework at length in The Economics of Infinite Desire.

Martin Heidegger, “The Question Concerning Technology” (1954). The concept of Bestand (standing-reserve) and the argument that the technological orientation reveals everything, including human beings, as resources held in readiness for further ordering.

Jonadas, The Friction That Was Thinking (2026). The spotter-vs-forklift distinction: the difference between a tool that removes difficulty and a tool that keeps difficulty on the person.

Jonadas, What Won’t Cross (2026). The formation-stage argument: the artifact crosses to the new medium, the person who was being formed by making it does not.

Jonadas, The Single-Player Game (2026). The argument that external scorekeeping quietly replaces the internal game, and that the switch does not announce itself.

Jonadas, The Economics of Simplified Living (2026). The argument that the gesture of refusal is itself available for mimetic imitation, and the case for honest examination over performative rejection.

The last paragraph had finally resolved. I had been sitting with it for forty minutes, the argument rearranging itself, the sentences not quite landing until, at some point I cannot pinpoint, they did. For a moment the file was just a file. The thinking was still warm in the room.

Then something shifted. Not a decision. A browser tab opening on its own. I was estimating. How would this land? Who would share it? Was the opening sharp enough for LinkedIn? The writing had stopped. I was scoring a performance.

I did not choose to switch games. The game switched, and I noticed it afterward, the way you notice you have been holding your breath only when you exhale.

Naval Ravikant has a name for what just happened. Modern life, he argues, runs on two fundamentally different kinds of games, and the most common mistake is applying the scoring logic of one to the other. The mistake is not dramatic. It is slow, directional, and by the time you notice it, the practice has already changed.

Two kinds of games

Most of what culture offers as success is a multiplayer competitive game: salary, status, professional recognition, follower counts. These games have external scores. Other people can see how you are doing. The leaderboard is real, the rules are legible, and winning means something, even if it does not, finally, satisfy.

But some activities have no valid external scoreboard. Peace. The quality of your attention. Character. The kind of clarity that comes from sustained, unperformed thinking. Naval puts it directly:

Warren Buffett makes the same distinction from a different angle. He asks: would you rather be known as the world’s best lover when you are actually the worst, or known as the world’s worst when you are actually the best? Most people give the approved answer. Most people live by the outer one.

Multiplayer games are real, and playing them well matters. Building a career is a multiplayer game. So is getting your work in front of the right people, or earning institutional standing for what you do. The problem is narrower: what happens when you apply multiplayer scoring to activities that have no valid external score. That application is a category error. And category errors, unlike honest mistakes, change the thing you are trying to measure.

The category error

A category error applies the wrong measurement framework to the wrong kind of object. The thermometer does not malfunction when you hold it to a piece of music; it asks a question the music cannot answer. Something analogous happens when you apply external scoring to a single-player game. The consequence goes further than imprecision: the measurement changes what is being measured.

Consider meditation. You can practice for presence or practice for a streak. The physical act is the same: ten minutes, eyes closed, attention returning to the breath. The orientation is not the same. One is a single-player game; the other has been converted into a multiplayer proxy. The measurement cannot capture what the practice is for, and it will increasingly govern the practice. Missing the streak becomes the failure. Presence becomes an unmeasured side effect of something the app is actually tracking.

Something begins as yours and gradually becomes a performance of itself.

Working through a problem on the page, building an argument, finding out what you believe by articulating it: this is the form of thinking I described in The Friction That Was Thinking as most at risk from frictionless tools. There are external signals that you are becoming clearer: a reader who follows the argument, a teacher who points out where you lost them. These signals have value. Making external validation the governing purpose is something else entirely.

When the essay is written toward engagement metrics, the prose starts to optimize for landing rather than for thinking. The opening gets punchy. Complexity gets smoothed. The sentence that is actually most precise gets cut because it will lose people. The subtler damage happens earlier, in the draft: the questions you choose to pursue are the questions that seem shareable. The writing that matters most, the thinking that does not yet know where it is going, stops happening, because it does not translate.

The apparatus is built for one kind of game. It makes the other kind harder, not by prohibition, but by making the multiplayer game constantly legible and the single-player game permanently silent. You are simply never shown the right scoreboard.

When the structure plays you

These activities are not purely internal. Thinking, teaching, producing knowledge: all have a genuinely social purpose. Teaching is for students. Knowledge is produced for a community. What external scores cannot capture is whether the people doing these things have internalized the standards that make the work worth doing, whether they care about what thinking is for, whether they teach in a way that transforms rather than merely delivers. A citation count may loosely track something. Governing your practice by it degrades the practice, in ways that are slow, cumulative, and very hard to see from inside.

I learned this in the academy. Academic departments are evaluated by exactly these metrics. The intention to resist them was genuine. It was also insufficient. The department’s standing, the funding available to colleagues, the institutional weight that protected junior researchers: all of it was downstream of multiplayer games I found philosophically illegible. Refusing to play did not mean accepting the consequences for myself. It meant accepting them for people who had not made that choice.

That is the bind. Individual integrity becomes insufficient when single-player activities are placed inside institutional structures that are themselves multiplayer games. The structure plays through everyone in it, regardless of their intentions. Opting out is not a personal sacrifice you are free to make; it is a decision with collective consequences you do not have the standing to impose on others.

The internal score

Think of a musician practicing alone. Not performing: practicing. A difficult passage, played again and again in an empty room. No audience. No recording. The standard is internal: whether the phrase lands the way it sounds in the mind, whether the hands do what they intend to do. The feedback is immediate and constitutive. The music either happens or it does not, and no amount of applause or indifference changes what happened in that room.

Now think about what changes when the same musician starts preparing primarily for the recording, the performance, the response. The technique may improve; recordings require precision. But the repertoire begins to drift toward what captures well, what has a clear narrative for the program notes, what audiences have heard before and will respond to with recognition. The difficult passage that reveals an unresolved gap in the playing gets abbreviated or avoided; it does not serve the performance. The practice room is still there. Something that used to happen in it is no longer happening.

This is not decline by any visible measure. The musician is performing more, building an audience, growing by every external indicator. The single-player game, the one that required the difficult passage, has been quietly replaced by a multiplayer proxy that looks identical from the outside and produces a different person on the inside.

Bernard Williams described this structure with some precision. Ground projects, he called them: commitments so constitutive of a person that their value cannot be evaluated from outside. They are not goals you hold the way you hold a quarterly objective. They are part of what makes you the particular person you are. Applying external scoring to them is a category error: the measurement does not merely fail to capture something real; it changes what is being valued, and therefore changes the person doing the valuing. You can ask whether the musician hit the notes. You cannot meaningfully ask whether the musician is becoming who they need to become. The question assumes a meeting point between the internal and the external that does not exist, roughly like asking what temperature jealousy is.

The single-player game has its own score. It is simply not available to anyone else.

The wrong scoreboard

The two categories can coexist. You can write for reach and for thinking, pursue recognition and genuine development, play the career game seriously and still know what you actually value. The practical question is narrower: which game is governing which practice, and do you know?

You cannot find out how you are doing at a single-player game by looking at a multiplayer scoreboard. The categories do not translate. And if you keep checking the wrong scoreboard long enough, you stop playing the right game. Not by decision. The game you were playing gradually becomes the game the scoreboard was measuring.

The musician does not decide to stop practicing the difficult passage. She decides to prepare for the performance, and the difficult passage quietly disappears from the program.

The meditator does not decide to stop practicing presence. He decides to protect the streak, and presence becomes something that may or may not be happening while the counter increments.

The writer does not decide to stop thinking on the page. She decides to check how the last piece landed, and the next draft begins a little closer to the scoreboard and a little further from the thought.

None of them notice the moment the game changes. The switch does not announce itself. It arrives as a reasonable question, an innocent glance at the dashboard, a harmless check. And then the practice is answering a different question than the one it started with.

I finished the draft. The argument had resolved. For a moment, the thinking was still warm in the room.

Then I opened the browser.

---

Sources & further reading

Naval Ravikant, in The Almanack of Naval Ravikant, ed. Eric Jorgenson (2020). The single-player/multiplayer distinction and the inner scorecard; pp. 143–144.

Bernard Williams, “Persons, Character and Morality,” in Moral Luck (1981). The account of ground projects and their constitutive relationship to personal identity.

Søren Kierkegaard, Concluding Unscientific Postscript (1846). The argument that subjective truth cannot be externally verified without being transformed into something other than itself.

Today, May 1, is Labor Day in much of the world. It feels like the right day to publish a piece I have been struggling with for a while, about what work might still be, once we let the machines have the parts they were always going to take.

It is almost three in the afternoon. You are about to approve a request that has crossed your screen a thousand times before. The fields are complete. The pattern matches. The customer is authorized, the dollar amount is below your threshold, the workflow has done its job. Your finger is already moving toward the approval button when, for half a second, it stops.

Nothing dramatic has happened. You did not catch an error. You did not remember a rule. You hesitated because something in the request, the timing, the wording, the order of the line items, the sender’s tone in the comment field, resembled another request that ended badly. You will not remember which one. You will probably not even know, in the next minute, that you stopped. But the hand knew before the mind.

Most office work is full of these tiny recognitions. The body knows where the friction is before the report does. The eye sees the dropdown and remembers the exception. The hand refuses, for half a second, before any reason has been produced. None of this feels dramatic from inside the day. It feels like getting through it.

We have also learned to perform some of this. The mouse jiggle to stay “active” on a chat app. The careful keyword in a status update so the pipeline reads it as priority. The way a ticket is phrased so the queue routes it past the colleague known to slow it down. We have already begun to format our presence for the systems that watch.

What changes now is that the format has a new reader.

Last week, TechCrunch reported, following Reuters, that Meta plans to capture employee mouse movements, clicks, keystrokes, and navigation patterns to train AI agents that will perform the same kinds of computer work the employees are doing. A few months earlier, OpenAI began rolling out memory features that, with the user’s permission, capture the rhythms of a person’s day across applications to give the model a longer continuity. The first records the work life. The second records the rest of life. The first is corporate; the second is voluntary. The difference between what the company coerces and what the user accepts is becoming difficult to feel.

The immediate scandal is surveillance. It should be. But there is something colder underneath the privacy problem.

The project makes sense.

If work has become a sequence of movements across interfaces, then record the movements. If the task lives in the path from ticket to spreadsheet to dashboard to approval flow, then capture the path. If a job can be shown as a choreography of clicks, pauses, fields, and confirmations, then the worker becomes, at least at that level, a training interface. The employee is not replaced after the work has been understood. The employee is used to define what counts as understanding the work.

That is the more disturbing thought. AI is not making us machine-readable from the outside. It is finding the parts of us that have already been rewritten for machines.

Operational legibility

The most disturbing thing about machine-readable work is not the recording. It is what the recording teaches us about how to appear. The person learns, slowly and without any single moment of consent, to present herself first in the form a machine can process. A biography becomes a data structure wearing a human name. The work trains a metaphysics.

A person becomes machine-readable when the important parts of her work can be captured as operational traces. Not the work as she experiences it. Not the responsibility she bears. Not the years that make one pause different from another. The trace: move here, click this, reject that, approve this, escalate there, leave behind a sequence another system can replay.

This is the next step beyond what I called The Writing That Was Never Yours. In that essay, I examined role-prose: the language we speak as a function, the mask we wear to be heard by the institution. Role-prose is what we say to be legible to other humans inside the system. Operational legibility is deeper. It is not what we say, but the residue of what we do. It is the click, the pause, the sequence of movements captured without our conscious authorship. Role-prose is the mask; operational legibility is the fingerprint left on the glass.

Elizabeth Anscombe, in her short and exact book on intention, made a distinction the office never made and that the AI age is forcing us to remember. The same physical movement, she argued, can be many actions, depending on the description under which the action falls. A man pumping water can be poisoning the household, supplying the vegetable garden, exercising his arm, or finishing his shift. Which of these the action is depends on the description that the agent can own. The log captures the gesture. The description is what has an owner.

Operational legibility erases the description. It preserves sequence and discards the reason the sequence was worth performing.

This did not begin with AI, or even with computers. Frederick Winslow Taylor’s scientific management, in 1911, made the worker visible as motion, timing, output, and deviation from the prescribed method. The body entered management as something to be decomposed. Later, the office learned to do the same to attention. Email timestamps, ticket queues, CRM fields, calendar density, productivity dashboards, version histories, response-time metrics: each made some part of professional life easier to count and compare.

Sometimes this was liberating. A system that depends entirely on someone’s private memory breaks when that person leaves. A bureaucracy, at its best, protects us from favoritism. A record can make power answerable. A process can keep a claim from vanishing because the wrong person was absent that day.

Legibility is not the enemy.

The danger begins when the trace stops serving the work and starts replacing our sense of what the work is. The dashboard does not show what happened. It shows what the dashboard can register. The ticket does not contain the case. It contains the case in the form the queue can route. The keystroke log does not capture a worker thinking. It captures the shadow thinking leaves when it passes through software.

That shadow is useful. It is also not the person.

Meta’s project is a nearly perfect image of the conversion. The employee is not asked to describe the work in language, with reasons, doubts, and responsibilities. The employee is asked to keep working while the system learns from the residue. The model does not need to know what the pause meant if it can learn where pauses usually occur. It does not need the lived weight of the decision if it can reproduce the sequence by which the decision normally appears.

Operational legibility is the condition in which the residue becomes enough.

The resume was already a parser

You can feel the same conversion in the modern job search.

A resume once promised, at least ideally, to present a life to a reader: not the whole life, of course, but a shaped account of a professional path. It had omissions, emphasis, tact. It asked someone to see a pattern in the years.

Then the reader became a parser.

Applicants learned the new etiquette quickly. Use standard headings. Remove unusual formatting. Repeat the exact keywords from the job description. Do not describe the work in the terms that are most accurate if the system is looking for a different phrase.

Do not be interesting before you are ingestible.

The two words name two registers, and the order is the whole problem. Interesting is the register a human uses for another human: a turn of phrase, a strange path, a shape of attention. Ingestible is the register a human uses for a parser: a pattern that survives tokenization, a keyword density, a header the software has been trained to recognize. The career advice, harmless on its surface, is metaphysics in disguise. It says: the parser comes first. Be readable, then, if there is room left, be a person.

There is rarely room left.

No one experiences this as philosophy. It is advice passed from recruiter to candidate, from career coach to student, from anxious applicant to anxious applicant. The metaphysics arrives without anyone having authored it.

The same training happens on platforms. A creator begins with a voice, which is to say a way of standing in relation to an audience. Then the graph begins to teach. This opening holds retention. This sentence loses the room. This face travels. This duration dies. The creator adjusts, then adjusts again, and slowly the work begins to arrive pre-shaped for distribution.

In machine learning, reinforcement learning from human feedback trains a model by rewarding some outputs and discouraging others. On platforms, the direction reverses. The algorithm trains the human. A person performs reinforcement learning on herself until the voice becomes easier for the system to carry.

What looks like self-expression is often already self-translation.

The resume and the creator feed are earlier versions of the Meta story. First, a person learns to present herself as a readable object. Then a machine becomes good at reading that object. Then we call the machine uncanny, as if the uncanniness began with it.

The manager as router

Inside organizations, the conversion is quieter because it looks like competence.

The manager opens the dashboard and sees the red items first. A stale ticket. A customer sentiment score. A delayed handoff. An owner missing from a field. A graph moving in the wrong direction. The world arrives already sorted by the system’s sense of salience.

The dashboard offers the relief of reduced complexity. By turning a messy human situation into a glowing red alert, it desaturates the weight of judgment. The dashboard is a filter that makes action possible while making judgment optional.

But the interface has a way of becoming the room.

After a while, the manager knows the work through the traces it leaves: the escalation, the field, the metric, the update. The habit becomes bodily. Eyes go first to the red number. Fingers move toward the routing action. The question “what happened here?” is quietly replaced by “what state should this object move to next?”

This is where the manager becomes legible as a router.

Some routing should be automated. There is no dignity in moving a ticket from one queue to another when the reason is already settled. Let the machine do that. The problem is that many organizations have made it hard to tell which moments were merely procedural and which only looked procedural because nobody had time to understand them.

An AI agent arrives, and the question becomes painfully reasonable: if the job is moving traces across interfaces, why does a person need to be there?

The answer cannot be that a person is always needed. That would be false. The answer has to be more exact. A person is needed where the trace is insufficient: where the case changes meaning when held against a history no field contains, where the metric improved because the underlying practice degraded, where the correct action would be wrong because this customer, this colleague, this student, this patient is not a token of the usual type.

The manager’s real work begins where routing fails. But if the institution only sees routing, it will train the manager to become what it can see.

The old name for this condition

Heidegger helps because he does not begin with the gadget. He begins with a way the world is allowed to appear.

In “The Question Concerning Technology,” he argues that modern technology is not merely a collection of instruments. It is a mode of revealing. Under Gestell, or enframing, things show up primarily as resources to be ordered, optimized, stored, and used. A river appears as hydroelectric capacity. A forest appears as timber inventory. A field appears as yield. The thing has not vanished. It has been made to appear in the aspect under which it can be mobilized.

That is why the Meta story feels less like a new scandal than a completed sentence. The worker appears as usable behavior. The day appears as training data. The pause appears as a statistical feature. The body at the computer appears as an input stream.

Heidegger’s word matters because “dehumanization” is too blunt. It makes the problem sound like a sentimental reduction of human richness by cold machines. The actual process is stranger. The human being remains present, active, clever, adaptive, overworked, sometimes even proud. But the system receives her under one aspect: as something whose usable pattern can be extracted.

That is enframing with a reader attached.

AI does not invent the frame. It reads what the frame has prepared. It is powerful exactly where the world has already been formatted as standing-reserve: the email, the status update, the keyworded resume, the optimized post, the ticket path, the dashboard action, the sequence of clicks through a business process. We keep asking whether the machine can imitate human work. A better question is how much of human work was already forced to imitate a machine-readable form.

The answer is uncomfortable because it is uneven. There are parts of our work we should be relieved to surrender. There are also parts we surrendered years ago while still performing them ourselves. The completed view is the precondition for a question that, until now, was not really available.

What the clearing reveals

Heidegger, near the end of the same essay, says something less quoted than Gestell and more important. The danger of enframing, he claims, is also the only place the saving power can show itself. “The danger, when grasped as the danger, becomes the saving power.” The phrasing sounds mystical and is not. He is saying that there is no escape from the danger to some unspoiled outside. There is only the difference between living inside the danger without seeing it and seeing it as the danger that it is. The seeing is already a turning.

What was making the seeing hard, all this time, was that we ourselves were performing the operational work. As long as the worker, the manager, the applicant, and the creator were the ones running on the dashboard, the dashboard could be confused for the work. We could pretend the metric, the parsable resume, the optimized post, and the ticket queue were what we did. We were the ones doing them, after all. We bled into them. They felt like work because they cost work.

AI changes that, not by replacing us but by separating us. When the agent runs on the dashboard, the dashboard stops being mistaken for the work. The keystroke log, captured by the model, is at last visible as keystrokes. The ticket path is at last visible as a path. The parsable resume is at last visible as a parse. We can finally see, because we are no longer the ones generating the residue, that the residue was never the building. It was scaffolding.

That is the saving power. Not a relief. An exposure.

The exposure forces a question that the era of the dashboard had quietly suppressed. If the operational was not the work, what was?

The question feels strange because the answer was always present and always demoted. The professional vocabulary of the late twentieth century filed it under headings like “soft skill,” “people side,” “judgment call,” “executive function,” “the human element.” The phrasing is telling. Soft, by contrast with the hard countable work. People side, as if the rest of the work were not. Judgment call, in the singular, as a moment of exception, rather than as a continuous capacity. The vocabulary made the actual work sound peripheral. The periphery sounded peripheral because the operational center had grown large enough to occupy the foreground.

The center was empty.

What was always there, behind the operational, is what every profession, in its older self-understanding, knew it was for. The doctor who has a sense of this patient, not patient-as-token. The teacher who hears that this student is asking a different question than the one the syllabus expects. The lawyer who recognizes that the case cannot be resolved by the precedent because the precedent assumed something that does not hold here. The manager who, from a way the meeting went silent, knows that the team has stopped trusting the project. The engineer who, looking at a build, feels that something is wrong before the test catches it. These are not exceptions to the work. They are the work, when the work is taken seriously. The operational was the way they got reported.

Acknowledgment as the form of work

Naming what was demoted requires a vocabulary older than management theory, kept alive by philosophers who refused the terms of the bureaucratic century. Stanley Cavell named it most carefully. He called it acknowledgment: the comportment in which one treats another person as that person, in their irreducible particularity, rather than as a token of a type. Acknowledgment is not interpretation; it is the stance from which interpretation can happen. It is not knowledge of the other; it is the openness in which the other can come to be known. It is what is missing from the most well-supported algorithmic decision and what is present, often without being noticed, in the briefest exchange between two people who actually see each other.

Acknowledgment is not the only thing the operational hid, but it is the one that gathers the others.

Around it sit judgment, in something like the Aristotelian sense of phronesis: the capacity to make a call you can stand behind, in a particular situation, when the rule does not by itself determine what to do. Not “decision under uncertainty as ranked by the model,” which is something a model can do well. The act of standing for the outcome.

Around it sits attention, in something like Simone Weil’s sense: looking at a situation until you actually see it, rather than classifying it until you can act on it. AI classifies endlessly. It does not, and cannot, attend, because attention requires that the attender be capable of being changed by what she attends to.

Around it sits responsibility, in the older sense in which someone’s name is anexed to a decision and they remain in a position to be addressed about it. Not the audit trail. The person who, when something goes wrong, can say I, and here is why, and here is what I knew, and here is what I would do differently.

These are not soft skills. They are the working content of being a person in a profession. The operational era did not invent them. It compressed them, demoted them, called them the human side, and made the rest the dashboard. AI, by absorbing the dashboard, can return them to the center, but only if we accept the surrender of the operational rather than competing with the machine on its ground.

The new measure of work is not how many cases you processed but how many you actually saw. Not how many tickets you closed but how many you closed in a way you could defend in front of a different person tomorrow. Not how many drafts you shipped but how many of them were yours under a description you can own. The shift is from quantity of trace to density of acknowledgment.

This is also why “perform humanity harder,” the obvious response to the Meta story, fails. The performance of authenticity is itself parsable. Platforms reward it; employers template it; the market turns it into a content strategy by lunch. Acknowledgment cannot be performed for the system because acknowledgment is not for the system. It is for the other person. Its illegibility is not a strategy; it is the form of the act.

Can Gelassenheit scale?

Heidegger had a name for the comportment that uses technology without being used by it. He called it Gelassenheit, usually translated as releasement: a yes-and-no said simultaneously to the technical world. Yes, we use the radio, the airplane, the calculator, the agent. We depend on them. We do not retreat into a Luddite refusal. And no, we do not let any of these devices tell us what our being is. We let the machine be a machine; we do not let it become our metaphysics. Heidegger paired this with what he called openness to the mystery: the recognition that even the technical world rests on a kind of disclosure that is not itself technical.

The yes-and-no is harder to live than it sounds. While the operational still flowed through human hands, it was always tempting to identify with what the hands were doing, because the hands were ours. Saying “yes” to the technical infrastructure and “no” to the metaphysical claim was a distinction without a clean cut. The infrastructure and the work were tangled. Now the cut becomes available. AI takes the operational; we can finally say yes to the operational as operational, without confusing it with the work, and no to the suggestion that the operational was ever the measure of who we were.

This is the saving power, said in the other vocabulary. The danger of enframing, when grasped as the danger, opens the clearing in which Gelassenheit becomes practicable. Not because Heidegger was naively right about technology, but because the consummation of Gestell is what makes its outline finally visible.

Heidegger himself did not believe this could scale. He imagined Gelassenheit as a small, contemplative, “here and now and in little things” practice, lived on country paths and in Black Forest stillness. He did not think a society could organize itself around it. He may have been wrong, or he may have failed at imagination. The reason to think he was wrong is precisely the one he could not have foreseen: the technological precondition for a Gelassenheit at scale did not exist for him. It begins to exist now.

A Gelassenheit at scale would not look like Gestell at scale. Gestell scales by uniformity, by making everything appear in the same aspect. Gelassenheit scales differently, by multiplication of singular acts. A society of acknowledgments scales the way a society of friendships scales: not by metric, but by texture. Many particular relations, each carrying its own weight, none reducible to a sum.

The popular version of this proposition is older than the AI debate. Star Trek has been running it for sixty years. The Federation handed the operational over to computers a long time ago. The replicator handles supply; the ship handles routing; the sensors handle parsing. Picard does not approve tickets. He deliberates about the Prime Directive. He sits with a grieving officer. He has a conversation with Data about what it means to be a person. The crew does the work that only crew can do: the call, the recognition, the attention to a particular situation in which a particular person is at stake. The starship is the precondition for the captain. The dashboard is what frees the deliberation, not what replaces it.

We mistake the show for utopia and miss that it is a coherent social proposition. A society in which most human work is non-operational is a society that has decided, deliberately, to let the operational be operational. It is not a fantasy of leisure. It is a fantasy of seriousness: the seriousness about each other that the operational era was too busy to maintain.

Whether such a society is achievable is a different question. There is nothing in the technology that requires us to reorganize this way. Capital can choose, and may choose, to keep humans inside the operational, racing against the machine. The platforms can choose, and have chosen, to monetize the performance of authenticity rather than the practice of acknowledgment. The danger does not save anyone automatically. It only gives us, for the first time, a clearing in which the work of saving could be done.

What the day already contains

Return to the scene at three in the afternoon.

You hesitated. You did not know why. The system, if it was watching, recorded the delay, time-stamped the next action, and moved on. Tomorrow it will use a few thousand such delays, summed across thousands of employees, to teach an agent how often the human at this stage of this workflow takes a beat. The agent will learn. It will not learn what the beat was for.

You knew. Or rather, your hand knew, and a part of you that does not separate from the hand knew with it. The request did not look like the others, in some way you could not list, and the not-looking-like was a piece of work you had spent years acquiring without noticing, and that work had a name. It was acknowledgment. The request was, briefly, a particular request, made by a particular person, against a history you held without naming. You looked. You did not classify. You attended.

The era of the dashboard called that the soft side of the job. The era of the AI exposes it as the side that is left when the dashboard goes to the machine. The hand pausing over the trackpad is not a vestige of an older way of working. It is the early form of a newer one, in which the operational has been delegated and the human is being asked to do, deliberately and at last, the thing that human work was always doing badly under the disguise of the operational.

The clearing is opening. It does not save us. It puts us in front of a question we had been allowed to avoid for a hundred years.

What is the work, if it is not what the dashboard could measure?

The day already has the answer in it. The pause was the work. The next call you make in which you actually see the person is the work. The decision you can stand behind, against the metric, is the work. The acknowledgment is the work. Heidegger’s “here and now and in little things” was never a smaller scale than the operational. It was the only scale at which acknowledgment is even possible. What changes now is that there may be room for many such here-and-nows, distributed, multiplied, taking up the foreground that the operational has finally vacated.

The question is whether we will claim the room.

---

Sources & further reading

TechCrunch / Reuters, “Meta will record employees’ keystrokes and use it to train its AI models” (2026). Reporting on Meta’s workplace data-capture initiative for training AI agents, including mouse movements, clicks, keystrokes, and navigation patterns.

OpenAI, “Pulse and continuity in personal AI” (2026). Public materials describing memory and continuity features that capture aspects of the user’s daily activity to inform the model across sessions.

Martin Heidegger, “The Question Concerning Technology” (1954). The central source for Gestell, or enframing, and for the formulation of the saving power that grows where the danger is.

Martin Heidegger, Discourse on Thinking (1959). The text in which Gelassenheit, releasement toward things and openness to the mystery, is laid out as the comportment that says yes-and-no to the technical world without retreating from it.

Frederick Winslow Taylor, The Principles of Scientific Management (1911). The early management dream of decomposing labor into measurable methods, timings, and outputs.

G. E. M. Anscombe, Intention (1957). The short and exact account of intentional action as action under a description that the agent can own.

Stanley Cavell, “Knowing and Acknowledging,” in Must We Mean What We Say? (1969), and The Claim of Reason (1979). The development of acknowledgment as a stance toward another person that is prior to interpretation, and that no body of evidence can substitute for.

Simone Weil, “Reflections on the Right Use of School Studies with a View to the Love of God” (1942). The classical account of attention as a moral act.

Maurice Merleau-Ponty, Phenomenology of Perception (1945). The background for the claim that skilled action is bodily contact with a meaningful world.

Jonadas, The Writing That Was Never Yours (2026). Companion essay on role-prose: professional writing that had already separated itself from authorship before AI arrived.

Jonadas, The Friction That Was Thinking (2026). Companion essay on the forms of resistance that build judgment rather than merely slowing output.

If you’ve ever felt an uncanny hollow while reading your own emails or status reports, it’s not because you’re a bad writer. It’s because the system wasn’t asking for a person. It was asking for an operation. This essay is a diagnostic for “Role-Prose”: the language of functions speaking to other functions, and why it is the first thing we should let the machine take.

In Portuguese, we called it prestação de contas: the rendering of accounts. In practice, it was an institutional proof of life. In the academy, it took the form of the research grant renewal or the annual activity report. In the corporate world, it is the quarterly review, the compliance audit, or the impact assessment. The names change, but the genre is identical: a thick series of templates demanding that you describe, in prescribed language, what you have accomplished to justify your funding, your time, or your existence.

Over more than twenty years, across fellowships, grant cycles, postdoctoral positions, and eleven years as a philosophy professor at a Brazilian federal university, I wrote versions of these documents. I also reviewed them, as a referee for funding agencies and peer-reviewed publications, occupying the other side of the same form. I was producer and reader at once, inside a closed system that everyone understood was not really communicating anything, and that everyone produced with the seriousness of someone who believed it was.

What struck me was not the tedium. Tedium I expected. What struck me was a feeling I could not name: I was writing, my hand was moving, words were appearing on the page, and I was not there. The philosophy conference that had changed how I thought about Wittgenstein became “capacity building” and “the dissemination of research outputs to relevant stakeholders.” The seminar where three students and I spent three weeks on a single argument in Kant became “pedagogical activities fostering critical analytical competence.” I was present in every formal sense. I was absent in the only sense that mattered.

The writing had my name on it. The writing was not mine.

This is the first kind of writing AI should take from us. Not because it is worthless. Not because it requires no intelligence. Because much of this writing had already separated itself from authorship before the machine arrived. AI did not dehumanize it. AI exposed that the dehumanization had already happened.

The uncomfortable truth is that a large part of professional writing was never the expression of a person. It was the speech of a role.

Role-Prose

Role-prose is what a function writes through a person.

The weekly status update. The grant accountability report. The meeting recap that exists so a meeting can prove it happened. The postmortem whose acceptable conclusion was visible before the investigation began. The executive summary of a document that was itself written because some other document had to be summarized upward.

From the register I spent twenty years inside: “With reference to the above-mentioned resolution, the undersigned hereby declares compliance with all applicable requirements.” “The research activities carried out during the reporting period are fully consistent with the objectives set forth in the approved project.” “Respectfully submitted for your consideration and deliberation.” No one says these things. The role says them.

In 2023, the cartoonist Tom Fishburne drew what may be the most efficient diagnosis of AI in the workplace. Two workers in adjacent cubicles. One uses AI to expand a bullet point into a professional email. The other uses AI to compress that email back into a bullet point. The loop is perfect. The meaning was always the bullet. The email was the institutional costume the message had to wear. What passed between the cubicles was never communication. It was a function’s artifact, inflated and deflated by whatever tools were available. The TL;DR became the TL;DW: Too Long; Didn’t Write.

This is not a satire of AI. It is a satire of what AI found waiting for it.

A practical test for role-prose: Would this sentence still work if anyone in the role had signed it? Was the acceptable conclusion visible before the writing began? Would anything be lost if no one had been personally present? If yes to all three, it is role-prose.

These texts are not useless. A status update can prevent confusion. A funding report can protect institutions from audit. A research brief can help a team avoid a stupid decision. The problem is not that role-prose has no value. The problem is that we kept confusing its value with authorship.

Role-prose wants competence. It wants accuracy, tone, format, institutional tact. It wants the sentence that sounds serious enough to be forwarded. It wants the conclusion stated firmly enough to guide action and softly enough to survive politics.

It does not want someone to appear.

Jean-Paul Sartre describes a waiter performing his role so completely that no person is visible behind it: movements too precise, manner too attentive, the person suppressed by the function. Role-prose is written in exactly that register. This is why AI imitation does not feel like imitation. There was never a person to imitate. The machine inhabited the function the person had been asked, over years, to become.

“For visibility.”

“As discussed.”

“We are aligned.”

“Further analysis is required.”

“The team is tracking this closely.”

“With reference to the above-mentioned.”

No one speaks this way because no one is speaking. The role is speaking. AI is good at role-prose because role-prose was already a machine language: the language of offices, templates, procedures, and people speaking as functions.

The Office Survives the Person

Max Weber saw the deeper structure before the inbox existed.

Bureaucracy was not, for Weber, merely red tape. It was one of the great technical inventions of modernity: offices, files, rules, written procedures, authority attached to position rather than personality. The point was continuity. The office survives the person. The file remembers what the body forgets. The role carries authority even when the individual changes.

This was not a pathology. It was an achievement.

A permit should not depend on whether the clerk likes you. A contract should not depend on whether the manager is feeling generous. Bureaucracy removes the arbitrary person so the system can become reliable. The impersonality of the form is one of civilization’s genuine protections.

But language changes when the person is removed.

In bureaucratic writing, the sentence is not primarily an expression. It is an operation. It records, authorizes, escalates, confirms, closes a loop. It does not ask to be memorable. It asks to be acceptable. It wants to survive audit, to be safely forwarded, to reduce ambiguity just enough for the next procedure to begin.

That is why role-prose has its distinctive texture: cautious, padded, abstract, mildly dead. Not bad because the people writing it are bad writers. Bloodless because the system is asking the blood to step aside.

Weber’s ideal was elegant in theory. It required conditions that do not exist everywhere. Brazil has a concept for this: the jeitinho brasileiro. The anthropologist Roberto DaMatta spent decades analyzing the tension between Brazil’s formal institutions, imported wholesale from European models, and the underlying social fabric, which has always prioritized personal relations. The rules exist. The impersonality was never delivered. The bloodlessness was required, and so was the blood.

The result is a double perversity: too impersonal where it matters to be addressed, too personal where it matters to be impersonal. At the bottom, the bloodless template nobody reads but everyone must produce. At the top, in institutions like the Supreme Court, the person erupts precisely where the impersonal was supposed to protect. A justice delivers an oral opinion lasting thirteen hours, or grounds a constitutional ruling in poetry. The theatrical personal performance appears exactly where the role’s restraint would have served justice better. Role-prose is what happens when a system can neither remove the person properly nor let the person appear at the right moment.

The Irreducible Claim

What remains?

Not everything with a template is merely role-prose. The same surface can hide different structures.

A research report is role-prose if it merely arranges known information into an expected format. It becomes something different the moment someone has to say, after sitting with the evidence: This is the risk that matters. A condolence note is addressed writing, written for a specific person whose situation changes every word. A draft that surprises the writer is discovery-writing: the author did not know what she thought until the document forced the conclusion into view.

The document type does not settle the question. The relation does.

Think of the last time you spent an hour calibrating the tone of a report: confident but not arrogant, humble but not weak, detailed but not exhausting. You were not writing. You were managing a function. The prose served the position you occupied, not the thought you had. Let AI do that. Verify it. Own the consequences. But do not confuse the removal of that labor with the loss of anything that mattered.

There is a different moment, when the work is real. The evidence does not tell you what to say. It only removes excuses. The data conflicts. The model is confident for reasons you do not trust. The politically convenient answer is available, and so is the evasive one.

Then someone has to decide what matters.

That moment is not role-prose. It is authorship under institutional pressure. The value of the report lives in the sentence someone is willing to stand behind. Not the sentence that summarizes the evidence. The sentence that risks being wrong.

AI can draft everything around that sentence. But if no one ever arrives at the claim they are willing to answer for, the report has no author. It has only a route.

This is why “AI wrote it” is too blunt. It can mean the machine handled the plumbing. It can mean the machine helped a person see more clearly. It can mean the machine supplied the judgment no one wanted to own. Only the last one is fatal.

The danger is not that AI will write too much. The danger is that we will stop asking what kind of writing we were doing.

Counterfeit Presence

Bureaucracy trained us to confuse signatures with presence.

The name on the document was the sign that someone was responsible. Often that was sufficient. The system did not need the full person: it needed a traceable role. But responsibility and presence are not the same thing.

You can be responsible for a document you did not author in the richer sense. Managers do this constantly. Executives do it with speeches. Modern institutions depend on this separation; Weber saw it as the secret of bureaucracy’s power. The office, because it was designed to make the person replaceable, outlasts any individual occupying it.

AI widens that separation. That may be the condition for doing less fake writing, for freeing people from hours spent laundering information into institutional tone, for exposing which documents were procedural all along.

But it also makes counterfeit presence cheap. A leader can send a warm note they did not write. A manager can deliver feedback they did not think through. A company can flood its people with language that sounds attentive while no one has attended. The surface improves as the relation thins.

That is the real danger: not that AI produces words without humans, but that it lets humans appear to have been present where they were only approving output.

The Writing That Is Yours

What authorship requires is not resistance, but presence.

Authentic writing has a phenomenological signature. The prose becomes transparent: you are not aware of choosing words, you feel the thought moving, resisting, turning back on itself. When a sentence will not close, it is not a grammar problem. It is a thinking problem. You have not yet understood what you are trying to say. The form is the pressure through which the thought becomes legible. The friction is the thought occurring.

The other experience is also recognizable. Sitting down to write what you already know you want to say, moving through the motions of a conclusion you arrived at before the work started. The words appear, the document fills, the form closes. And you have the uncanny sense that you could have been anyone. The function produced the document. You were the occasion.

You can feel the difference from the inside. The prose that runs in a groove cut before you sat down. The conclusion waiting at the end of the template. The words that come too easily because they are not really yours to choose. Something registers the absence: even when the document looks professional, even when it receives praise, even when it is technically correct in every particular.

Authorship means being present where the relation requires a person. Not just traceable. Present: someone stayed with the difficulty, changed their mind, arrived at a sentence they did not plan and are willing to answer for.

The accountability reports can go. The status updates can go. The compliance memos, the institutional-tone renderings of information that was already available, the executive summaries of documents that were themselves summaries: the role-prose can be handled by a role.

What cannot go is the writing where you do not yet know what you think. The writing addressed to a specific person whose situation changes every word. The writing where your name on the page is not a signature but an appearance: the sign that someone showed up, stayed with the difficulty, and arrived at something no formula could have guaranteed.

That writing was always yours. The machine cannot take it because the machine cannot do what produces it: be undecided, and present, and at risk.

The voice was already averaged in most of what we wrote at work. Let that go. What stands alone is smaller, clearer, and harder than we expected. It asks more of us than the comfortable mass of role-prose ever did.

---

Sources and Further Reading

Max Weber, “Bureaucracy” in Economy and Society (1921/1922). The office survives the person because the office was designed to make the person replaceable. The key claim this essay tests: that impersonality is a protection, not a pathology. Brazil complicates both sides.

Roberto DaMatta, O que faz o brasil, Brasil? (1984). DaMatta’s anthropological analysis of the tension between Brazil’s formal, universalist institutions and the personalist social fabric underlying them. The jeitinho brasileiro as structural consequence: the form demands impersonality; the reality delivers neither impersonality nor the protection it was supposed to provide.

Migalhas, “Fux vota durante 13 horas...” (2024). Reporting on Justice Luiz Fux’s marathon oral opinion at the Brazilian Supreme Court, illustrating the theatrical eruption of the person where the role’s restraint was expected.

AMB, “A poesia e a literatura nos ajudam a fugir da subinterpretação...” (2012). Former Justice Carlos Ayres Britto defending the structural use of literary and poetic rhetoric in constitutional reasoning.

Jean-Paul Sartre, Being and Nothingness (1943). The waiter in bad faith: movements too precise, manner too attentive, person suppressed by role. Role-prose is the linguistic form of bad faith, not a lie but a suppression.

Maurice Merleau-Ponty, Phenomenology of Perception (1945). The blind man’s cane: a tool that, with practice, becomes transparent, the point where the person feels rather than an object the person holds. The phenomenology of authentic writing draws on this: prose becomes the medium through which you feel for something not yet touched.

Hannah Arendt, The Human Condition (1958). The distinction between what a person is (function, role, category) and who appears through speech and action. Role-prose eliminates the “who” by design.

David Graeber, Bullshit Jobs: A Theory (2018). The accountability report as a species of bullshit work: not necessarily useless, but structured to eliminate the person performing it. Graeber’s account of the moral injury, producing institutional seriousness one does not believe in, is the most precise available.

Tom Fishburne (Marketoonist), “AI Written, AI Read” (2023). The cartoon this essay cannot improve on: one cubicle uses AI to expand a bullet point into an email; the adjacent cubicle uses AI to compress the email back into a bullet point. The message was always the bullet. The email was the institutional costume.

Jonadas, The Thought You Didn’t Have (2026). The companion essay on discovery-writing: what AI takes is not the finished thought but the process by which the thought would have been formed.

Jonadas, What Won’t Cross (2026). The formation stage as what does not survive automation.

A followup to What Won’t Cross. That piece made the general case: formation happens in friction, and the gym for thinking has to be built. This one walks inside and picks up the equipment.

The grain changed direction and the plane told me before my eye could.

I was finishing the face of a walnut panel with a hand plane. A plane wants to be pushed with the grain. When you find the direction, the shaving comes off in a single continuous curl, the surface is left glassy, and there is nothing in particular to think about. When the grain turns, the tool stops curling and starts tearing. The change arrived in my wrists first. A hesitation in the blade. A small sound that was not the right sound. The tearout was not yet visible on the face of the panel when I had already lifted the plane and flipped the board.

That lesson was not written anywhere. It was in the pushback.

Matthew Crawford, who left a think-tank job to open a motorcycle repair shop, observed that a repairman has to begin each job by getting outside his own head and noticing things: he has to look carefully and listen to the ailing machine. The machine is not cooperative. It does not return a clean answer to a clean query. You have to attend to it because it requires your attention — it will not reveal what it is doing to someone who is not fully present. It resists, and in the resistance it tells you what it is actually doing, which is almost never what the manual said it was doing.

The wood resists. The room you are teaching resists. The patient whose story does not fit the template resists. In every practice worth its name, something refuses to cooperate, and that refusal is not an obstacle to the work. It is the teacher.

I wrote about what formation is and why it does not cross from one medium to the next. That essay named the loss. This one is about what it would take to replace what was lost.

The loop without the middle

Every practice you care about moves through a cycle: a cue, a routine, a reward. You see a gap in the argument; you work out the position; it holds under scrutiny. You see a patient; you form an impression; it guides the treatment. The cue fires, the routine runs, the reward arrives.

But the routine is never only a behavior. It is the part of the work that forms you. Working out the position is where the thinker learns what she actually believes. Forming the impression is where the physician learns what the chart cannot carry. Something is deposited in the person doing the work: a sensitivity to cases of this kind, an instinct for when a case is not of this kind, a feel for where the work has to be held more carefully than its outputs require.

That deposit is the practice. The reward is just what tells you the loop closed.

Here is what is happening. A system is inserted between the cue and the reward, executing the routine for you. The gap is registered. The position arrives. The reward is collected. The loop closes faster than it ever has.

And nothing forms.

Think of the last time you felt a genuine shift in your own understanding. It did not happen when a perfectly formatted document arrived in your inbox. It happened when a paragraph you were writing refused to resolve, and you had to sit there for twenty minutes until you found the word that fixed it, or when a piece of code failed three times and you had to trace the logic back to the premise you didn’t know you had assumed. That frustration was not the cost of the thought. That frustration was the thought occurring.

The completed loop is what the habit theorists describe without knowing they are describing a problem. The cue fires, the reward arrives, and the middle, the part where resistance would have deposited something in you, has been bypassed. You will not notice it for a while. Skilled practitioners do not lose their skill overnight. The lawyer who has read ten thousand briefs does not forget how to read briefs when she stops writing them herself. The senior engineer who spent a decade debugging race conditions does not forget race conditions when the first pass comes from a model. The loss is downstream: the practice that does not form in the next generation, and the attunement that stops being refreshed in this one. I have called this specific liability calibration debt in another essay; here I want to examine what produces it, and what would prevent it.

What a gym actually is

A gym is a designed environment for the production of difficulty. Not for the output of difficulty, not for trophies or race times, but for difficulty itself, as the point. The gym produces no artifact. What it produces, over time, is the person. And it is engineered, deliberately, to prevent anyone from finding a shortcut.

Good gym design has four features. Equipment that is adjustable, so the load can match your current capacity and grow with it. A protected structure, so the session cannot be cancelled just because there are easier things to do. Other people who take it seriously, because the social environment either reinforces the practice or quietly erodes it. And someone who knows the difference between productive struggle and actual injury, who will let you sit with the hard thing for as long as it is building you and step in only at genuine failure.

We have not built this for thought.

What we have instead is an accelerating pressure to close every loop as fast as possible, because the closed loop is the deliverable, and every open loop looks, from the outside, like wasted time. The structures that used to protect cognitive time under load, the uninterrupted afternoon, the paper-only draft, the meeting where the senior actually read your memo, have been eliminated one by one, and nobody called it a loss because the loops kept closing and the rewards kept arriving.

These four gym features translate directly into design for knowledge work, and I will come back to each of them. But first it helps to understand, precisely, what the load is.

Load and duration

In strength training, time under tension refers to how long a muscle is under load during a set. A person who swings a barbell up with momentum and drops it has moved the weight, but the muscle was under tension for perhaps two seconds. A person who lifts the same weight slowly, pauses at the top, and lowers it under control has kept the muscle under tension for six or eight seconds. The second person lifted less impressively. The second person built more.

The completed loop is the cognitive equivalent of the momentum rep. The cue fires, the system swings the output into place, the loop closes. Time under cognitive tension: near zero.

This is not just an elegant analogy. The brain is physical tissue. Synaptic plasticity, the biological formation of new neural pathways, requires metabolic energy. It is expensive for the body. The resistance of a hard problem is the metabolic trigger that commands the tissue to adapt. When the machine removes the cognitive load, it is not just saving time. It is literally depriving the brain of the metabolic tension required to build a new circuit.

Designing for cognitive time under tension means holding the loop open deliberately. Not because you enjoy inefficiency. Because the time you spend sitting with the unresolved problem, with the argument that is not working, with the data that refuses your model, is the time in which your capacity is being physically built. The delay is not a bug. The open loop is where you were being made into someone who can handle the next version of this problem.

I notice this in my own practice. I read long books and take notes on them, revising the notes as I go. I listen to long-form podcasts, sometimes three or four hours, without skipping, without “multi-tasking”. Sometimes, I will run the material through a quick summary pass first, the kind an AI can generate in seconds, to know whether the conversation is worth the investment before I make it. But the summary is a filter, not a substitute. Once I decide the material is worth the full engagement, I give it the full engagement, every digression and tangent, every place where the argument stalls and recovers. The summary told me whether to sit down to the meal. The summary is not the meal. What the full engagement deposits is not the information, which the summary could have delivered. It is the cognitive paths that duration forces me down, the thoughts I arrive at only because I was required to follow the argument long enough to find them. I could not have reached those thoughts on purpose. I could only have been led there by the material, over time, under sustained load.

That is deliberate cognitive time under tension. The summary pass is the momentum rep. The full engagement is the slow, controlled rep: the mind stays under load, the resistance persists, and something is deposited that the summary would not have deposited.

The difficulty curve

The body adapts. If you lift the same weight every week, the stimulus stops producing growth. The load has to increase. Trainers call this progressive overload: the deliberate, incremental raising of difficulty to keep pace with developing capacity.

AI, as currently deployed, does the opposite. It flattens the difficulty curve. The hard argument becomes as tractable as the routine one. The complex analysis becomes as easy as the simple one. The junior professional’s cognitive load stays constant regardless of how demanding the material actually is, because the system absorbs the surplus difficulty before she encounters it. This is the equivalent of going to the gym every day and lifting the same five-pound weight for years.

A 2025 study of thoracic oncology tumor boards found that AI recommendations agreed with the human clinical team seventy-six percent of the time. The twenty-four percent where they disagreed is the part of the work where clinical judgment still lives: a radiologist remembering a case that ended badly, a surgeon sensing that the imaging does not tell the whole story, the pause where someone asks a question the data would not have prompted. If the junior clinician is trained on a workflow in which the summary is already written, the likely diagnosis is already named, and the first-pass analysis is already done, she will never encounter the load of the twenty-four percent. She will lift the same weight for her entire residency. The seniors will carry the hard cases until the seniors retire. Then there will be physicians who were never overloaded, and who therefore never grew.

The same logic runs through any practice where difficulty is not uniformly distributed. Legal analysis. Code review. Strategic synthesis. The hard cases are where the load actually is. A workflow that smooths difficulty to a constant floor does not produce experts. It produces practitioners who are permanently competent at the mode and permanently fragile at the tails, which is precisely where the work matters most.

The result is a condition you can already feel inside organizations: phantom competency. A junior worker whose outputs are flawless but who feels an intense, quiet impostor syndrome because she knows her capacity has not grown. She is shipping senior-level work while feeling like an intern, acutely aware that she could not reproduce her own deliverables if the network went down. The organization measures the artifact and reports success. The person measures her own formation and feels hollow.

The forklift and the spotter

We have been treating AI as a forklift. Its job is to move the cognitive freight: take the rough argument, deliver the polished version; take the raw data, deliver the summary; take the intake, deliver the differential. The forklift does not care about the person standing next to it. It cares about the load.

A gym does not use forklifts. A gym uses spotters.

A spotter stands behind the bench while you press. The spotter does not lift the weight. The spotter watches you struggle with it, lets you sit in the tension for as long as the tension is productive, and touches the bar only when you reach genuine failure. Even then, the spotter applies the minimum force needed to keep the bar moving. The entire point is that you take the weight. The spotter’s job is to make sure the weight does not kill you, not to spare you from it.

The difference between a forklift and a spotter is the difference between two relationships to difficulty. The forklift assumes difficulty is waste. The spotter assumes difficulty is the product. One optimizes for the artifact. The other optimizes for the person.

Almost everything in the current conversation about AI in professional life is organized around the forklift model. How do we move the cognitive load faster, with fewer errors, at lower cost? That question is not wrong. There is an enormous quantity of cognitive freight that should be moved by machine: formatting, scheduling, boilerplate, summaries of meetings you were not in. Let the forklift take it.

The problem is that the forklift does not know the difference between freight and exercise. It picks up everything. And once it picks up the exercise, the person standing next to it gets weaker without noticing, because the loop still closes, the deliverables still ship, and the reward still arrives.

Consider what happens when a system architect reaches for a whiteboard at the hardest moments of a project. The whiteboard is slower than the screen. The marker cannot undo. There is no copy-paste. She is not choosing it because she enjoys the smell of dry-erase. She is choosing it because the constraints of the medium force her to synthesize before she executes. The inability to undo is the load. The slowness is the tension. She is acting as her own spotter: using the minimum level of tool support that keeps the weight on her, and reaching for the more capable tool only when she genuinely cannot proceed without it.

That is what the spotter model looks like when it is working. Not refusing the machine. Not performing analog virtue. Choosing, deliberately, the level of support that keeps the difficulty on the person.

What the gym requires

Back to the four features of good gym design, and what they mean for a knowledge worker.

• Adjustable load: The work a person encounters must include difficulty proportional to their developing capacity, not smoothed to a constant floor by automated assistance. The twenty-four percent of hard clinical cases is not a problem to be resolved before it reaches the trainee. It is the training. Routing hard cases around junior practitioners in the name of efficiency is the organizational equivalent of removing all the heavy weights from the gym and replacing them with five-pound dumbbells, forever.

• Protected time: The cognitive session cannot be cancelled because there are easier things to do. Unassisted drafts. Paper-only reviews. Meetings where the primary documents were read by humans, not summarized by machines. These are artificial constraints in exactly the same way the barbell is artificial: they exist to produce a condition that the environment no longer generates on its own. The cognitive gym is engineered resistance. It produces nothing except the person.

• Serious community: The social environment reinforces the practice rather than the shortcut. If everyone around you submits AI-drafted work without resistance, the person who drafts her own has no signal that the practice is worth the time. Institutions that want formation have to make the practice legible and reward the attempt, not only the output, because the output is the one thing the machine can also produce.

• A knowledgeable spotter: Someone who can calibrate exactly how much help to withhold, and for how long, before the load becomes damage rather than development. This is not the same as a mentor who gives advice. It is the rarer thing: someone whose own formation is intact enough that they can distinguish productive struggle from waste, and who has the discipline not to intervene too soon. In The Thought You Didn’t Have, I traced this in the specific case of writing: the senior who lets the junior’s draft stay clumsy long enough for the junior to find out what she thinks is doing something that the senior who improves the draft immediately is not doing.

None of these are anti-technology propositions. The gym is not against transport. A cognitive gym is not against AI. It is a designed environment for a specific kind of formation that automation does not produce and cannot substitute, in the same way that a running track is a designed environment for a kind of fitness that the car does not produce and cannot substitute.

The choice being made right now

Every organization deploying AI at scale is currently making a choice about this, whether or not it has named the choice. It can deploy AI as a forklift, moving every available load, optimizing every loop, closing every gap between need and output. It will have faster, smoother, cheaper production for as long as the people running it still carry inherited capacity. Or it can deploy AI as a spotter: present at every point of genuine failure, absent at every point of productive struggle, calibrated to keep the weight on the person for as long as that is where the building happens.

The second option is harder to design, harder to measure, and harder to defend in a quarterly review. It also produces people whose judgment can be trusted when the situation has no precedent and the machine has no useful answer. That situation is arriving faster than the forklift model can handle.

The wood taught me by resisting the blade. The long book teaches me by refusing to summarize itself. The long essay teaches me by refusing to resolve in the first three sections. These are not inefficiencies. They are reps.

For any practice you care about, the question is where the resistance still is, and who is meeting it. If the difficulty is still there and you are still the one encountering it, the practice is forming you. If the difficulty has been absorbed and something else is meeting it on your behalf, the loop is completing without you. There are practices that should complete without you. Formatting should. Boilerplate should. Summaries of meetings you were not in should. But the practices whose resistance was building you are your weights. The inconvenience is the load.

The question is not whether to use the forklift. You will. So will everyone around you. The question is whether anyone has decided which weights are yours to carry, and whether the organization you work inside has any interest in protecting that decision, or whether it is, right now, quietly picking up everything it can reach.

---

Sources & further reading

Matthew B. Crawford, Shop Class as Soulcraft (2009). The argument that skilled manual work is not a lower form of cognition but its own form of it. The description of what attention the ailing machine compels is the sentence this essay opens on.

Jonadas, What Won’t Cross (2026). The companion essay that names the formation stage as the part of professional work that does not survive automation. This essay assumes that argument and asks the design follow-up.

Jonadas, Calibration Debt (2026). The argument that firms were accidental calibration apparatuses, and that AI is dismantling the apparatus while retaining the revenue engine. The junior-professional and senior-executive patterns here draw on that framework.

Jonadas, The Thought You Didn’t Have (2026). The specific case of writing as formation practice: what AI quietly takes is not the voice but the process by which the writer was being made into someone with a voice.

K. Anders Ericsson et al., “The Role of Deliberate Practice in the Acquisition of Expert Performance,” Psychological Review (1993). The research behind the popular “10,000 hours” formulation. Ericsson’s actual argument, which the popular version dropped, was that the key variable was deliberate practice under feedback and resistance, not hours alone. Time under tension, not time spent.

Hubert Dreyfus and Sean Dorrance Kelly, All Things Shining (2011). On poiesis, the craftsman’s way of bringing things out at their best. Dreyfus’s wider work on skill acquisition and embodied expertise stands behind the completed-loop argument here.

Journal of Clinical Medicine, Large Language Models in Multidisciplinary Tumor Boards (2025). The seventy-six percent agreement figure between AI recommendations and human clinical teams. Source

We are very good at recognizing when our bodies need resistance to stay healthy. We invent entire industries to put that resistance back into our physical lives. We are surprisingly bad at recognizing when our minds need exactly the same thing. Here is what happens when we automate the friction away.

Twenty years ago I had a supervisor who wrote the first drafts of his papers by hand, late into his career, on yellow legal pads, forming block letters one at a time like a printing press. He had a computer. He used it for everything else. But the first draft, the one that would eventually become an article, he wrote on paper with a pen. I remember thinking it was eccentric. He was not old, not precious, not performing anything. He just refused to skip that stage.

It took me about fifteen years, and the arrival of a drafting assistant that could turn my rough sentences into clean prose inside of a minute, to understand what he had been protecting.

He was guarding something that, at the time, I could not name.

Nobody argues against exercise. We all agree the body needs it. What we sometimes forget is how recent the agreement is, or rather, how recent the problem is. For most of human history, physical labor was built into survival. You did not need a gym because the field, the workshop, the docks kept you strong as a byproduct of staying alive. The modern gym exists because industrialization quietly removed that labor and left the calories. The treadmill goes nowhere. The barbell serves no purpose beyond the act of lifting it. The gym is engineered resistance, designed to replace the resistance that civilization accidentally took away.

Something similar is now happening to thought. And we have not built the replacement.

The tools of cognitive production (AI assistants, synthesis engines, summarizers, drafting agents) have become so fluent that the friction which once accompanied knowledge work is vanishing. That friction was never only an obstacle to the output. It was the condition under which the person doing the work was being formed. Remove it, and the artifacts keep arriving. What stops arriving, quietly and without anyone noticing for a while, is the capacity of the person behind them.

What is vanishing is the formation stage of the work: the stretch of the process where the person making the artifact is also being made by it. The draft that taught you what you thought. The code that taught you how the system actually behaved. The diagnosis that taught you what the patient had. These are not production stages. They are not recording stages. They are the hours during which, while producing an artifact, a specific capacity is being built inside the person producing it.

The formation stage is invisible in the finished work. A reader of the polished essay cannot see it. A reader of a fully AI-generated essay cannot see that it is missing either. The difference between the two versions is not on the page. It is in whether someone became someone by making one of them.

A confession

I say this as the last person who should be arguing it. I was an early adopter from the moment I had access. I started in academia in the nineties and I used computers for everything I could put them near. I hated writing by hand so much that when I had to take notes I abandoned cursive and formed block letters one at a time. Like a printing press. It felt more natural. It still does.

When I began teaching in a philosophy department and brought a laptop to the first seminar and used PowerPoint for my lecture, I could feel the disapproval. Philosophy, the implicit argument went, was not something one did with slides. It took over a decade for the department to stop minding. By the time I left, almost everyone was using them.

So when I say the thing I am saying, understand that I am not defending any medium against any other. I have never been nostalgic for the older tool. What I am reaching for is not about the tool. It is about whether, somewhere inside the work, a person is still doing the thinking.

Here is how I write now. I open a blank page and type (sometimes, dictate) a first version in my own voice. The sentences are rough. I know they will not stay this way. An assistant sits beside me that can clean them up within minutes once I hand them over. I could skip the rough version entirely. It would be faster. The prose would read better. I do not do it, because if I did, I would not have thought anything.

I wrote about this in The Thought You Didn’t Have: sometimes a sentence resists me, and I rewrite it four times, and somewhere in the fifth rewrite the shape changes, and what was wrong with it turns out to be that I had not yet understood what I was trying to say. The sentence was waiting for me to catch up. That kind of discovery cannot be outsourced. Once you hand over the drafting, the thought that would have arrived through the resistance simply never arrives. You get something that sounds like the thought. You do not get the thought.

My supervisor was protecting the formation stage of his own work. He did not say “I think with the pen.” He just would not skip to the clean version.

Which stage makes you someone

Before you decide where you stand on any of this, run a different test. Pick the kind of work you are most worried AI is threatening. Writing. Teaching. Coding. Medicine. Lawyering. Strategy. Pick one. Now make an honest list of every electronic aid you already use without a second thought:

Spell-check. Grammar check. Word processors with auto-format. Search engines. Autocomplete. Stack Overflow. Email filters and auto-reply suggestions. Document templates. Citation managers. Version control. Slide decks generated from bullet points. Transcription software. Calendars that schedule your meetings for you. Dashboards that summarize what happened. Tools that condense what others said.

Every item on that list is something you already handed to a machine, some of them decades ago. If any of those things had been the work, you would not have handed them over. You did. So they were not the work. Cross them out.

Spell-check. Grammar check. Word processors with auto-format. Search engines. Autocomplete. Stack Overflow. Email filters and auto-reply suggestions. Document templates. Citation managers. Version control. Slide decks generated from bullet points. Transcription software. Calendars that schedule your meetings for you. Dashboards that summarize what happened. Tools that condense what others said.

What remains?

What remains, for most people who do this honestly, is smaller than they expected and more specific than they can quickly name. It is whatever stage of the process is the one where they actually struggle. The research direction they followed without knowing where it was going. The first draft of an argument that surprised them. The moment they had to rethink their entire framing because the evidence would not fit. The piece of code whose right shape became visible only after four wrong shapes. The conversation with a specific student where they found themselves saying something they had not said before.

Those stages are not defined by medium or tool. They are defined by the presence of resistance: the moment the material did not do what you wanted, and you had to meet it, and find something on the other side. The meeting is what thinking is. The artifact is only partial evidence that someone was once there doing it.

The test becomes this: if I could remove this stage from my workflow and replace it with a machine that delivered the artifact cleanly, would I still be the same person on the other end? If yes, let the machine do it. If no, that is the formation stage of the work, and it is not for sale.

What must go unassisted

In January 2026, Tyler Cowen gave a lecture at the University of Austin (full lecture; the section on writing and thought begins at 21:50). Cowen is an economist, a prolific blogger at Marginal Revolution, and one of the most forceful advocates for artificial intelligence in American intellectual life. He uses AI reading agents to filter his own incoming text. He has written, publicly, that he composes partly for the machine indexes that will process his work, a logic I examined in The Claim Upon the Training Data. He is not a nostalgic.

His position on AI in education is therefore worth attending to, because he is not defending a boundary for sentimental reasons. And the position is two-handed.

Cowen argued that the institutional response to AI has been organized almost entirely around the wrong problem: how to prevent students from using tools that are freely available and professionally essential. This, he suggested, is an evasion. The real question is how to restructure education when AI can produce many of its traditional outputs faster and better than students can.

Alongside any expansive integration, Cowen insists that students must spend real unassisted hours writing, no AI access, because writing is how thinking gets done, not how it gets recorded. You do not have a thought and then write it down. You have the thought by writing, against the resistance of the page and your own unformed attempt. Take the resistance away and the thinking was skipped.

The argument is not about handwriting. Cowen was not proposing fountain pens. The practice being protected is medium-neutral: laptop, phone, whatever. What is protected is a portion of the student’s time during which the student is actually meeting their own cognitive resistance and producing, alone, the rough version whose function is not the paper that comes out of it but the person who comes out of it. Without those hours, the four years produce a file of polished papers and no interior.

That is a formation stage protected as policy. It is one of the only places in a current curriculum where that protection is being proposed explicitly. It has also been proposed before, in circumstances we now find slightly ridiculous.

What the press could not copy

In 1492, a Benedictine abbot named Johannes Trithemius, at the monastery of Sponheim in the Rhineland, wrote a small treatise called De Laude Scriptorum, “In Praise of Scribes.” The printing press had been in commercial operation for four decades. The scriptorium was collapsing across Europe. Most of the scribes Trithemius had known in his youth were out of work. The demand for hand-copied texts was drying up, the economics were finished, and the great Benedictine scribal tradition that Trithemius himself had spent his life inside was, in the working sense, already over.

He wrote the treatise as a defense. Forming each letter, he argued, was a form of contemplative attention. Parchment outlasts paper. The scribe who copies a text has read it in a way the printer has not. A monk who spends twenty years in a scriptorium is not the same creature as a man who operates a press for twenty years.

In 1494, he had the treatise printed.

He had to. The press was already the only serious distribution network for new ideas in the German-speaking lands. An argument published only in manuscript form in 1494 would reach a few dozen monastic libraries and die there. If he wanted the argument to live at all, he had to print it. He chose reach, and in choosing reach he conceded, structurally, the surface of the argument he was trying to make.

Trithemius lost the argument he thought he was making. The tract defending scribal labor survives today because a press copied it. The specific craft he was defending, the particular room of men and parchment and candlelight, disappeared within a generation. On the surface, he was wrong, and history agreed.

Underneath the argument about the medium was another argument he did not quite manage to articulate, and on that one he was right.

The monk who spent twenty years copying sacred texts was not producing copies as a side effect of his day. He was being formed into a certain kind of reader and a certain kind of person by the labor. The slow forming of each letter bound his attention to the text. The years of shared work in the scriptorium bound him to the men beside him. Specific capacities were being built in him: a sense for when a scribe ahead of him had drifted, when a passage had been garbled in an earlier copy, when a doctrinal turn hinged on a single preposition. None of this was about the parchment. It was about the hours.

The press took the production. That was fine, the press did it better. What the press could not do was the hours. Those hours were never a production stage. They were a formation stage: the place where a mind met the resistance of a text across years, and a specific kind of person was made from that meeting. The monastery ended up with more books than the scribes had ever produced. It ran out of the specific kind of person that twenty years of copying used to form.

The artifact crosses. It always crosses. Put it on parchment, move it to print, scan it to PDF, ingest it into a model, it crosses. The formation does not cross. It has to be done, in place, by someone, against resistance, for the time it takes.

Formation, up the ladder

The same question, in the same shape, applies at every altitude where judgment is made.

Consider the teacher on the other side of the desk. A professor who reads a student’s paper carefully, with this specific student in mind, thinking about how to help this particular person improve, is doing the teaching. Whether the professor reads on paper or on screen does not matter. What matters is whether the professor is meeting the resistance of this student’s specific move, this student’s specific misunderstanding, this student’s history across the semester. A professor who hands the paper to an AI grader and forwards the output has not taught anything that day. Not because the comments are wrong. The comments may be better. Nothing has been taught because nothing has been met.

This is not a medium distinction. The same AI tool, used with full engagement or as a way to skip the engagement, produces two different outcomes. I watched the same split happen in my own department with PowerPoint when it (finally) became ubiquitous. Some colleagues took slide preparation as an occasion to rethink the lecture from scratch. Some pasted yesterday’s bullets and read them aloud. Same tool. Two different acts. The first group was still teaching. The second group had stopped.

The same collapse is now playing out in corporate life, and I traced the economic logic of it in Calibration Debt. The law firm, the consultancy, the hospital: these institutions were, accidentally and as a side effect of proximity, apprenticeship apparatuses. A junior analyst sitting next to a senior analyst was not only producing deliverables. She was being calibrated, developing the feel for when an argument overshoots its evidence, when a model is cheating, when the client’s stated problem is not the actual problem. That calibration was never designed. It was the residue of shared, uncomfortable work done in proximity over years.

When a senior partner delegates the reading of a junior’s brief to an AI, receives the summary, and ships revised clean prose, the brief improves, the mentorship vanishes, and the junior accumulates what I called calibration debt: the liability that grows whenever someone deploys AI to produce work beyond their ability to evaluate it. The brief files perfectly. The formation does not happen.

This extends all the way up, with one feature that makes the top of the ladder more exposed than the bottom.

Picture a CEO running the same exposure at a higher altitude. A board deck arrives on Monday. The assistant summarizes it. The model drafts a response. She edits lightly and ships. Over a quarter, the practice that used to keep her strategic judgment calibrated, reading the raw document, forming a first-pass view, catching the anomaly that the summary would have flattened, has been retired. She will not notice the retirement. Nothing will announce it. The briefings keep arriving. The decisions keep shipping. The deliverables look, if anything, sharper than before.

The difference is that the junior at least has someone above her who might catch the gap. The CEO does not. There is no senior partner reviewing the CEO’s thinking. When the capacity to evaluate one’s own strategic synthesis atrophies, and it will, because the practice that maintained it has been retired, the error will not be caught by anyone in the room, because the room was built on the assumption that the person at the top had done the thinking.

Formation debt at the top of an organization is a risk that has no current name and no current check. The gym equivalent is easy to picture: an executive who believes himself strong because the people around him all lift on his behalf.

The absent sender

South Park caught one side of this earlier and sharper than most philosophy has. In “Deep Learning” (Season 26, Episode 4, March 2023), Stan begins using ChatGPT to write his texts to Wendy. The texts are immediately better, more thoughtful, specific, attuned. Wendy, feeling outmatched, starts using ChatGPT to respond. The exchange improves in surface quality continuously. Neither of them is writing anything. Neither of them is thinking about the other. Two phones exchange perfect text between ghosts.

When Wendy finds out, what she is shocked by is not that Stan’s texts were bad. They were the best she had ever received from him. What she is shocked by is that she had been in conversation with no one.

This is the formation argument turned outward. Formation is the private half: what the person becomes by doing the work. Attendance is the public half: what the other side recognizes in the artifact as evidence that someone sat with the material, with them specifically in mind. Both are residues of the same hours. When AI mediates the whole exchange, neither residue appears. No one is being formed. No one is being attended to.

An institution in which professors generate AI feedback on AI-generated papers, a university in which neither side of the exchange is actually encountering the other, has not become more productive. It has become a system of artifacts passing between parties who are no longer thinking about each other. The quality of the surface masks the absence of what was supposed to happen underneath.

What human communication is always reaching for, across every medium and every century, is one specific signal: did someone think about me. Not content. Content can be generated. The receiver is looking for evidence of acknowledgment. Handwritten letters survived typewriters by decades, not because handwriting is noble (I have already confessed I do not think it is), but because a handwritten letter was, for a while, the most available proof that someone had sat down for some number of minutes and attended to the reader specifically. When the signal became fakeable, people reached for whatever the next medium could not yet fake: the specific detail, the voice note, the message that could only have come from someone who actually knows. The form keeps changing. The thing being looked for stays constant.

There is more to say about what this does to communication itself, to the shape of friendship and mentorship and public life under an infrastructure that makes ghostwriting instant and invisible at every turn. That is a different essay. For now, the point inside this one is smaller and harder: the collapse is symmetric. What the sender does not do, the receiver cannot find.

The gym for thinking

Here is where the analogy circles back, and where it goes further than you might expect.

The Greeks did not separate physical and intellectual training. A gymnasium was part athletic facility, part philosophical seminar. Plato held his discussions in one, which is where the word Academy comes from: a public garden adjacent to the wrestling grounds of Athens, named after the hero Academus. The logic was the same for both: capability is not a faculty you are born with. It is a capacity built through structured practice under pressure, in a specific kind of space, with others who take it seriously. The exercise produces the exerciser.

We inherited the physical gym. We seem to have forgotten the thinking one.

What Cowen is proposing, unassisted writing alongside maximum AI integration, is the beginning of rebuilding it. Not a sentimental exception. Not an anti-technology gesture. A designed and protected practice of deliberate cognitive resistance, in the same way the gym is a designed and protected practice of deliberate physical resistance.

The same principle applies wherever judgment matters. For students: protected hours of writing as the formation practice. For junior professionals: apprenticeship structures that require actually encountering and being corrected by the work, not only receiving the improved output. For senior executives: a discipline of reading raw signals, forming a first-pass synthesis with their own minds, actually sitting with the material before the machine prepares its briefing. None of this is a call to inefficiency. The gym does not make you less mobile. It makes you capable of being more so. Protected thinking practice does not slow down an organization. It produces people whose judgment can actually be trusted when speed is required.

We need to build these rooms. We need to name them. We need to protect them from the entirely reasonable pressure to optimize them away, because the pressure to optimize them will sound, in every instance, like an argument for making things better and faster, and it will be right about the output and wrong about the person.

My supervisor with the yellow pad was guarding a formation stage. He just refused to skip it. He was right about something that the rest of us, now, have to make explicit and deliberate, because the accident that used to keep us in contact with our own resistance has been retired.

The artifact will cross in any medium. The formation will not. It has to be done, in place, by someone, meeting resistance, for the time it takes. What we have to decide, this decade, is whether to build the rooms for it on purpose, or to discover, once the need has vanished from everyone’s calendar and no one can remember what it produced, that we have lost the capacity to recognize what we lost.

---

Sources & further reading

Tyler Cowen, lecture at the University of Austin (January 8, 2026). Full video; the section on writing and thought begins at 21:50; Q&A on cognitive atrophy from AI dependency at 47:56. Cowen’s broader commentary on AI and education runs continuously through Marginal Revolution (marginalrevolution.com).

Jonadas, The Thought You Didn’t Have (2026). The argument that AI does not take the thought, it takes the process by which the thought would have been formed.

Jonadas, The Claim Upon the Training Data (2026). The essay examining Cowen’s practice of writing for machine indexes, and the argument that institutional founding happens at the fiction layer, not in the text itself.

Jonadas, Calibration Debt (2026). The argument that firms were accidental calibration apparatuses, and that the AI transition is dismantling the apparatus while retaining the revenue engine. The law firm and executive examples in this essay draw on Calibration Debt‘s central argument.

Johannes Trithemius, In Praise of Scribes / De Laude Scriptorum (1492, printed 1494). Translated by Roland Behrendt, edited by Klaus Arnold.

South Park, “Deep Learning” (Season 26, Episode 4, March 8, 2023). Written by Trey Parker.

Elizabeth L. Eisenstein, The Printing Press as an Agent of Change (1979). The field-defining account of what the press actually disrupted.

K. Anders Ericsson et al., The Role of Deliberate Practice in the Acquisition of Expert Performance (1993). The research behind the popular “10,000 hours” rule. Ericsson himself argued, repeatedly, that the key variable was not hours but deliberate practice under feedback and resistance, a nuance the popular version dropped and this essay relies on.

Stanley Cavell, The Claim of Reason (1979). The philosophical background for the distinction between having the words of another and actually meeting them.