When AI Makes the Ads, Who Becomes the Art Director?

The paid work that taught Monet to see

Claude Monet was born in Paris on November 14, 1840. In 1845, his family moved to the port city of Le Havre. By 1855, when Monet was fourteen, the town knew him for caricatures: quick, exaggerated portraits of recognizable local people. The surviving drawings are dated 1855 and 1856. No diary gives us the day when a harbor regular first handed the young Monet money, but the record is clear about the role the work played. These caricatures were probably his first paid artworks. 31

The commissions were small commercial pictures, not the beginning of a masterpiece. Monet drew the brokers and ship captains who frequented Le Havre's harbor cafes with oversized heads. He bought vellum, pencils, and penknives at Eugène Boudin's stationer's shop. The shop framed the drawings and put them in its window to sell. That window mattered twice: it gave Monet a small market and gave Boudin, already a landscape painter, a reason to notice the teenager's eye. 26 31

The meeting is dated to 1856, when Monet was fifteen. Boudin did not hand him a finished style called Impressionism. He invited him to work outdoors on the Normandy coast. The lesson was an encounter with a changing world: water, weather, air, and light kept altering the scene before the painter's eyes. A brushstroke was no longer just a mark that resembled an old picture. It was a decision made under conditions that pushed back. 26 27

By August and September 1858, Monet was showing a landscape of Rouelles, a place Boudin had taken him, at Le Havre's Société des Amis des Arts. The caricaturist had not simply acquired a prettier output. He had acquired a different practice: go to a place, expose an intuition to what is there, make a mark, look again, and revise. That loop is where technique becomes judgment. 31

Eighteen years later, on April 15, 1874, Monet and other independent artists opened their first exhibition in the photographer Nadar's Paris studio. Monet's Impression, Sunrise, painted in 1872, appeared there. A hostile critic used its title to mock the group as "impressionists." The insult stuck. But the style did not emerge from Monet alone, or from a brush alone. It came from mentors, peers, places, materials, exhibition choices, and years of paid and unpaid looking. 27 32

This is the point that a model's surprising combinations can obscure. It did not make Impressionism by learning what a wet brush, a canvas, the coastal air, and a changing sky could do together. It learned the visible residue called "Impressionism" only after painters had lived through that practice and left paintings, photographs, and reproductions behind. If no one had made those paintings, there would be no Impressionism in the archive for a model to retrieve, interpolate, or imitate. 30

Style is more than an arrangement of pixels or an inventory of ideas. It is a way of making that has been discovered in contact with a medium. Paint thins or clots. A brush drags, skips, or splinters. Pigment sinks into one canvas and sits on another. A scene changes while the painter is still deciding what to do. Those encounters can create a practice, a new set of questions, and eventually a new genre. A model can make an image that looks as if it belongs to that genre. It does not thereby make the practice that gave the genre a future.

Imagine a new kind of paint, one that spreads through canvas differently in dry air and damp air, or catches the light in a way no existing pigment does. Before artists have spilled it, ruined canvases with it, learned how to control it, and made work that other artists answer, there is no visual language for a model to learn. It may generate an attractive guess at what such art could look like, because it can borrow from existing styles. But the paint's actual possibilities arrive only through new material encounters. Once people have done that work, the resulting images and techniques can be captured and recombined. The archive follows the creation. It does not substitute for the event that created it. 30

This is not an argument that a computer can never be connected to cameras, tools, experiments, or human critique. It is an argument about the ordinary systems now buying up commercial image work. In a 2024 survey, 26% of responding illustrators said they had lost work to generative AI. In 2025, the Interactive Advertising Bureau reported that half of advertisers were already using generative AI to build video ads, and buyers expected it to power 40% of video ads in 2026. These figures do not measure how many future Monets, designers, or art directors have lost a mentor. They do show that the paid work through which a young person gets repetitions, feedback, and a living is changing first. The danger is not only fewer jobs. It is that we keep extracting the record of earlier human creation while weakening the ordinary paid practice that produces the next record. 28 29

The counterfeit is good enough, and that is exactly the problem

Let me concede the strongest form of the opposing case immediately, because everything I want to say depends on it being true.

For an enormous range of real work, AI output is fine. Not fine as a euphemism. Genuinely, adequately, appropriately fine.

A product brochure for a small manufacturer. A serviceable logo for a food truck. A landing page. A stock illustration for a slide deck nobody will look at twice. A first-pass summary of a document. Boilerplate contract language. A blog post whose only job is to exist. Routine code that has been written ten thousand times before.

These tasks share a structure, and it is precisely the structure the first essay identified: they are dense with data, conventional in form, low in stakes, and their failure modes are visible. They live deep inside the recorded space. There is nothing to discover, nothing to verify that a non-expert cannot verify by looking. The machine interpolates, and interpolation is the correct operation, and the result is good.

If this were the whole story, the story would be a happy one. A useful tool has made routine work cheap. That has happened many times, and on net it has usually been good.

But there is a second fact, and the collision between the two facts is the subject of this essay.

The second fact is that the routine tier was never merely routine. It was performing three functions that had nothing to do with the brochure.

It was the income that kept practitioners alive while they got deep.
It was the training ground where they got deep.
And it was the filter through which the public learned, dimly, what depth even looks like.

The brochure paid for the paintings. The junior analyst’s tedious data cleaning was the education that produced the senior analyst. The commercial illustrator who spent Tuesday drawing a storyboard for an ad agency spent Wednesday on the work that mattered — and the Tuesday work was not a distraction from the Wednesday work. It was the Wednesday work’s funding, and it was also its practice.

We are now removing the routine tier. And we are removing it under the impression that we are removing only a brochure.

Akerlof: the market for lemons in expert work

Imagine you need a used car next week. You have a new job across town, the bus takes ninety minutes, and the cars in your price range all look reassuringly similar online. Each ad has clean photographs. Each seller says the vehicle has been well maintained. Each car starts when you turn the key. One has a transmission that will fail within six months. One has been serviced carefully for years and will run reliably for a decade. From the curb, they are nearly identical.

You can inspect the paint, take a test drive, and pay a mechanic for an opinion. None of that lets you see the next ten years. You know that some fraction of the cars on the lot are bad, and you know that a seller will describe a bad car as a good one if the description costs nothing. So you do the sensible thing: you offer a price that reflects the average car you expect to find, not the best car the seller says is there.

Now switch places. You own the well-maintained car. You know the transmission was rebuilt, the oil changed on time, and the expensive repair that will not be needed next winter. The buyer cannot see any of it. She offers the average price, the price that makes sense for a lot containing both good cars and lemons. It is below what your car is worth to you, so you keep the car.

The owner of the lemon does not face the same choice. The average price is generous for a car with a hidden failure. They sell. Next month, buyers return to a lot with fewer good cars, but they cannot observe that change directly. They lower the price they are willing to offer. Another tier of good-car owners decides not to sell. The lot gets worse, the price falls again, and the market begins to empty of the very cars buyers hoped to find.

There is an economic model for what happens when buyers cannot tell quality apart, and it won a Nobel Prize.

Akerlof’s insight was that information asymmetry does not merely make markets inefficient. It can destroy them entirely — it can make good products unsellable at any price, not because nobody wants them but because nobody can identify them. 1

Now hold that next to the argument of the first essay.

The first essay’s central claim was that verification is the scarce good. That generation has become free and judgment has not. That the person without depth cannot tell a brilliant output from a catastrophically wrong one, because they look identical.

That is not merely an epistemic problem. Stated in Akerlof’s language, it is a market-destroying condition.

A client commissioning a brand identity cannot, in most cases, distinguish a design produced by someone with fifteen years of typographic judgment from one produced by a person who typed four sentences into a model. Both arrive as a PDF. Both look professional. The difference — the reason one will still work in five years and across eleven contexts and the other will fall apart the moment it meets a real constraint — is invisible at the moment of purchase. It is only visible later, and often never, and by then the invoice is paid.

So the client pays the average price. And the average price is now being dragged toward the marginal cost of a prompt, which is approximately zero.

And the people with fifteen years of typographic judgment withdraw. Not because they lost a competition on quality. Because they lost a competition on price, in a market that could not perceive quality.

The artists have a word for this, and they arrived at it before the economists did. They call it “good enough.” In their own accounts, they are consistently clear that they are not worried the machine is better than they are. They are worried that clients, managers, and consumers will deem it passable — and that being deemed passable is sufficient to end their ability to earn a living.

That is Akerlof, described from inside the lemon market by the good cars.

Gresham’s Law says the same thing in a sentence: bad money drives out good — but only when the two are forced to trade at par. The whole mechanism hinges on the counterfeit being accepted as equivalent. And that acceptance is not a fact about the counterfeit. It is a fact about the buyer’s inability to tell.

Just keep a reader on hand

There is a natural objection to the first essay, and it is worth taking seriously because it sounds like a refutation. It goes like this:

You made too much of the reader. Yes — the monkey cannot know it typed Hamlet. But so what? We do not need every monkey to be Shakespeare. We only need one good reader in the room. Let the machines and the amateurs generate a billion pages; as long as we keep a few real experts on hand to verify, the good pages will be found. Generation is democratised, verification stays with the experts, and everyone wins. You have not shown that the amateur must become an expert — you have shown that we must keep some experts around to check. That is a much smaller problem.

This is the right objection, and answering it precisely is the hinge of this entire essay. Because the answer is: verification is not a layer. It is a resource. And resources run out.

The Erdős story from the first essay is usually told as a triumph, and it is one — but look at what it actually required. An amateur produced a candidate proof. That candidate did not become knowledge by existing. It became knowledge by travelling: to an online forum, through people who understood it well enough to take it seriously, up a chain of increasingly expert readers, and finally to Terence Tao — one of the few humans on earth able to say not just “this is correct” but “this matters, and here is why.” The verification worked because the pipeline had spare capacity, and because the volume was small enough that a candidate of genuine merit could rise through it and reach the one reader who counted.

Now change one variable. Keep the readers exactly as expert as before. Just multiply the pile.

What happens to verification when generation becomes free is not that it gets harder. It is that it becomes impossible to afford — because every generated candidate imposes a verification cost on a human expert whose time did not become free, and the ratio of real-to-worthless collapses precisely because generating the worthless became costless.

This already happened, and we can name the date

We do not have to argue this hypothetically, because a clean, documented instance has already run to completion — in the one domain, software security, where verification is supposed to be cheap.

The curl project — a piece of open-source infrastructure running on billions of devices — had operated a bug-bounty program since 2019, paying researchers to find real security vulnerabilities. It was a functioning verification market: reporters submit candidate flaws, maintainers verify them, genuine findings get paid. Over its life it paid out more than $90,000 across 81 legitimate awards.

In January 2026, curl’s maintainers shut the program down. Not because it ran out of money. Because generative AI had made it trivial to produce plausible-looking vulnerability reports at zero cost, and the maintainers — a tiny team of human experts — were drowning. Daniel Stenberg, curl’s lead, described the project as effectively being DDoSed, and said that if he could, he would charge the submitters for the time they were wasting. By mid-2025 only about 5% of submissions were genuine vulnerabilities, while roughly 20% were identifiable AI slop — reports that, in one maintainer’s description, look legitimate at first glance and therefore have to be read before they can be refuted. That last clause is the whole problem. A worthless report is not free to reject. Someone qualified has to spend the time to know it is worthless, and that time is the scarce resource the flood consumes. 2 3

And curl is not an outlier — it is the visible edge of a pattern. The Python Software Foundation’s security lead flagged the same wave hitting CPython, pip, urllib3, and Requests. CycloneDX abandoned its bug-bounty program entirely. Across HackerOne, by some accounts 60–80% of submissions became invalid. The platform is being asked to verify more and more candidates, of which a smaller and smaller fraction are real — which is to say, the cost of finding each genuine finding is climbing toward infinity.

This is the counterargument’s fatal flaw, demonstrated in the wild. Keeping a good reader on hand does not save you if you cannot get the reader to the page that matters — and you cannot, when ten thousand confident, plausible, worthless pages arrive for every real one, and each must be read to be dismissed. The reader does not need to be fooled. The reader only needs to be buried.

And note the second turn of the screw, because it connects directly to the rest of this essay. A maintainer at another project, describing how they might cope, proposed restricting submissions to verified researchers only — and immediately noted that doing so would make it harder for junior researchers to break into the field in the first place. The defence against the flood is itself another cut to the pipeline that produces the next generation of readers. The counterfeit does not only bury the expert. It walls off the path by which someone might have become one.

The cost gradient: why this is a mild warning, not the disaster

Here is the part that should genuinely frighten you, and it is the reason this essay exists.

Software security is the cheap case. Verifying a vulnerability report is, relatively speaking, fast: a competent human can often read it, attempt to reproduce it, and know within minutes to hours whether it is real. Mathematics is similar — the verifier is a proof assistant like Lean, and a chain of expert readers, and while the reading is slow, it is tractable and it is cheap in resources: a laptop and a mind. These are the domains with the cheapest verifiers on earth, and even they are buckling under AI-generated volume.

Now walk up the cost gradient.

The first essay established that AI advances where a cheap external verifier exists — Lean for math, compilers for code, the Protein Data Bank for structure. What it did not dwell on is what “cheap” is hiding, and what happens to the argument when the verifier is not cheap. Consider the full ladder of what it costs to check a candidate:

Now run the counterargument on cancer.

“A bright amateur plus AI will generate a cure; we just need experts to verify the good candidate.” Set aside — as the first essay already argued — that generating the plausible molecule was never the hard part. Grant, for the sake of argument, that the AI hands you a genuinely promising candidate. Who verifies it, and with what?

Verification here is not a mathematician reading a proof over a weekend. It is target validation in a wet lab, against a literature where only ~11% of landmark preclinical findings replicate. It is synthesis, assays, pharmacokinetics, toxicology. It is animal models and then, eventually, a clinical trial that takes thirteen years and costs on the order of a billion dollars, and fails 97% of the time. That is the cost of verifying one candidate. And a cure is never one candidate — it is thousands of iterations, each requiring the same machinery, each needing an expert to decide which direction to modify next, because the data is far too sparse for any gradient to point the way. 4 5

The amateur with a chatbot can generate ten thousand cancer-drug hypotheses before breakfast. The world can verify a few dozen candidates per decade. The bottleneck is not the idea. It never was. The bottleneck is a verification capacity that is fixed, expensive, human, and already fully subscribed — and generation just went to infinity while it stayed flat.

This is why “keep a reader on hand” fails catastrophically outside the cheap-verifier domains. In math and code, the flood merely overwhelms the readers — bad enough, as curl shows. In biomedicine, materials, energy, the physical sciences — the places where the real stakes and the real innovations are — there is no cheap reader to overwhelm in the first place. There is only a slow, costly, physical process of verification that cannot be scaled by adding GPUs, and cannot be shortcut by anyone, expert or not, because the world itself is the verifier and the world takes its time.

And notice the deeper thing this reveals about what verification is in most of the domains that matter. Outside the formal corner — outside proofs and code, where a symbolic oracle can pass judgment — verification is not an act of reading. It is an act of building. To verify a business idea you do not inspect the pitch deck; you build the product, ship it, and test it against a real market, which takes years and usually says no. To verify a scientific hypothesis you do not admire the reasoning; you build the lab, run the experiment, and wait for the physical world to answer. The candidate and its verification are separated by an enormous, irreducible construction cost — capital, equipment, people, and above all time — and that cost does not fall when generation becomes free. It stays exactly where it was. So the free flood of plausible candidates arrives at a gate that still has to be opened one candidate at a time, by hand, at full price. AI can write a thousand business plans and a thousand grant proposals in an afternoon; it cannot build a thousand companies or run a thousand trials, and neither can the amateur holding the prompt. The generation is weightless. The verification has mass. And it is the mass that was always the real work.

Take a drug candidate all the way to a prescription

A candidate therapy has to survive target validation, synthesis, assays, pharmacokinetics, toxicology, animal work, three phases of clinical trials, manufacturing, and regulatory review. Wong, Siah, and Lo estimate that 13.8% of drug-development programs entering Phase I eventually receive approval. In a separate sample of drugs approved from 2009 to 2018, Wouters, McKee, and Luyten estimate median capitalized research-and-development investment of $1.14 billion per approved medicine, including the cost of failed trials. 4 23

Those are not merely accounting figures. They are the price of asking reality questions that no model can settle from text alone: does the molecule reach the tissue, help the patient, avoid a toxicity signal, and still work when the biology becomes heterogeneous? The reproducibility literature is a reminder that an early result is a reason to investigate, not a free oracle. Generating ten thousand hypotheses does not generate the laboratories, patients, time, and expert judgment required to decide which, if any, should become knowledge. 5

Take a power-plant concept all the way to the grid

The same constraint appears in energy. A model can propose a reactor geometry, a control policy, or a new materials stack in an afternoon. Electricity arrives only after a physical system can be fabricated, sited, permitted, fueled, commissioned, run safely, and connected to the network. For scale, the EIA's 2025 base estimate for an advanced-nuclear plant is $5,530 per kilowatt in 2023 dollars. At that rate, one gigawatt of capacity is roughly $5.53 billion in overnight capital cost; the EIA's reference unit is 2.234 gigawatts, or about $12.35 billion before financing costs and the cost of operating the plant. 24

Building the plant is still not the same as delivering power. A project must enter an interconnection queue, complete impact and facilities studies, fund any required network upgrades, sign an interconnection agreement, synchronize to the system, and reach commercial operation. Lawrence Berkeley National Laboratory reports that, for regions with available data, projects built in 2025 took more than five years at the median from their interconnection request to commercial operation. At the end of 2025, 2,061 gigawatts of proposed generation and storage were actively seeking connection in U.S. queues. 25

The reason experts remain indispensable there is not merely that they can verify. It is that they are the only ones who can decide which handful of the ten thousand candidates is even worth spending the verification budget on — which is the problem-selection judgement the first essay called taste, and it is the single most valuable act in any resource-constrained science, because it is the act that determines whether the decade and the billion dollars are spent on something real. The amateur cannot do it, and the AI cannot do it, for the same reason established at length already: there is no oracle, the data is sparse, the landscape is jagged, and only a mind that has lived in the domain can read a discrepancy as promising rather than merely possible.

Deep expertise, as this series keeps insisting, is necessary and not sufficient for the creative leap. But for the humbler, prior task of pointing the scarce verification budget at the right candidate — it is simply necessary, with no qualification at all.

That is why the answer is not to keep one expert on hand as a final stamp. Before a clinical trial, pilot plant, or grid connection can consume the scarce physical verifier, a proposal has to pass through many kinds of expert judgment: people who can identify a nonstarter, test whether the mechanism is coherent, expose the missing measurement, compare the claim with prior failures, design the next experiment, and decide whether the expected value justifies the next allocation of money, facilities, and time. 4 23 24 25

One person cannot do all of that, and no responsible institution should ask them to. The reader is a distributed system: domain specialists, experimentalists, engineers, reviewers, operators, regulators, and decision-makers, each able to reject some bad candidates before they absorb the next expensive stage. Their shared job is not to guarantee that every selected proposal will succeed. It is to make the pipeline selective enough that the few proposals sent to the physical world are meaningful bets rather than merely plausible pages.

The low-hanging fruit was the ladder

Here is where the economics stops being an inconvenience and starts being a catastrophe.

The market is not just squeezing artists. It is squeezing juniors.

Look at what the data actually shows.

Demirci, Hannane and Zhu studied online freelance markets around the release of ChatGPT and the image models. Within eight months of ChatGPT, postings for automation-prone freelance work fell about 21% relative to manual-intensive work. Writing fell 30%. After the release of the image generators, postings for graphic design fell about 18%, and 3D modelling about 16%. A study in Organization Science found freelancers in more AI-exposed occupations saw contracts fall around 2% and earnings fall about 5%. 7 8

Now look at where the damage is concentrated, because this is the part that matters.

Brynjolfsson, Chandar and Chen, using ADP payroll records covering millions of American workers, found what they called canaries in the coal mine:

• Early-career workers (ages 22–25) in AI-exposed occupations saw a 16% relative employment decline.
• For 22–25-year-old software developers specifically: roughly a 20% decline since late 2022.
• Employment for experienced workers in the same occupations remained stable — and for workers aged 35–49 it grew over 9%.
• The adjustment came through employment, not wages — firms stopped hiring rather than cutting pay.
• And the declines were concentrated in occupations where AI automates rather than augments.

Read those five findings together and a shape appears that ought to alarm everyone.

The bottom rung of the ladder is being sawn off. The top of the ladder is fine.

And UNESCO’s 2026 survey of the creative industries found the identical pattern, reported by the people living it: emerging artists are struggling more than established ones. Senior freelancers have relationships, reputations, and clients who know what they are buying. The junior has none of that, and competes with the machine directly, on the exact tier of work the machine does well.

(In fairness, this is contested. Some analyses — Gimbel and colleagues, and work from the Economic Innovation Group — find no clear break in aggregate employment trends. The aggregate may well be fine. But the aggregate is not the thing I am worried about. A pipeline can be fatally severed while the headline numbers look healthy for a decade, because the people already on the ladder are still standing on it. The damage from a broken pipeline does not show up in the year you break it. It shows up in twenty.)

The cross-subsidy nobody costed

Every practitioner I have ever read about had a day job inside the domain.

Bach was a church employee grinding out cantatas on deadline. Every novelist teaches, or edits, or writes copy. The illustrator quoted in one of these surveys spent twenty years in comics and publishing, but made most of his income drawing internal storyboards for advertising agencies — work that was never published, never seen, and never mattered to anyone. He describes it, touchingly, as still being paid to do the thing he loved most.

That work vanished, in his account, essentially overnight in 2023.

We keep describing this as artists losing gigs. It is not. It is the removal of the economic substrate on which the entire practice of art rests. The storyboard was not the art. The storyboard was the scaffolding that held the artist in position long enough to make the art — and, not incidentally, the daily drawing practice that kept the hand alive.

Sherwin Rosen described the economics of superstars: when a technology allows one performer to serve everyone, the middle of the distribution collapses. A handful of names capture the market, and everyone else is left with nothing. What AI has done to creative labour is a superstar dynamic with the machine itself in the superstar position — and the middle of that distribution is not a luxury. It is where the field reproduces itself.

A field with only stars and machines has no next generation of stars, because nobody survives the decade it takes to become one.

The irony, at civilisational scale

The first essay introduced Lisanne Bainbridge’s Ironies of Automation: the more reliable the automation, the less the operator practises, the more their skill decays — and yet the operator is retained precisely to handle the cases the automation cannot. 9

We can now state the economic version, which is worse:

The tasks AI does well are the tasks juniors used to do.
Doing those tasks is how juniors became seniors.
And seniors are the only people who can verify what the AI produces.

So we are automating away the manufacturing process for the only people capable of checking the machine.

This is not a metaphor and it is not a slippery slope. It is a supply chain, and we have severed it at the input end while continuing to draw down the inventory at the output end. The inventory is the current generation of experts. They are middle-aged. They will retire.

The classroom: how to counterfeit a semester

The market kills the apprenticeship from one side. The school kills it from the other — and here the damage is subtler, because it is entirely self-inflicted, and it feels like learning.

The difficulty was the point

There is a body of research in cognitive psychology that should be required reading for every educator alive right now, and it is called desirable difficulties.

Robert and Elizabeth Bjork established that conditions which make learning feel harder and slower often produce better long-term retention and transfer — while conditions that make it feel smooth and easy often produce fluent performance now and nothing at all in six months. Spacing beats massing. Interleaving beats blocking. Testing yourself beats re-reading. The generation effect: you remember what you produced, not what you read. Roediger and Karpicke’s work on retrieval practice: the act of dragging something out of memory is what strengthens it; passively reviewing it does almost nothing. 11 12

The unifying principle is brutal and counterintuitive: the feeling of ease is not a signal that learning is happening. It is frequently a signal that it is not.

Now consider what a large language model is, in a classroom, from a student’s point of view.

It is a machine for the total elimination of desirable difficulty.

The struggle is the mechanism. The struggle is not a regrettable cost of learning that we have finally found a way to optimise away. The struggle is the thing that does the learning. And we have handed every student on earth a device that removes it, at no cost, invisibly, with a result that looks exactly like the result of having done the work.

The scaffold that never fades

John Sweller’s cognitive load theory offers a genuine complication, and I want to be fair to it, because it is the strongest thing the optimists have.

Sweller showed that novices learn better from worked examples than from unguided problem-solving. Being told the solution path, at first, is better than flailing. So guidance is good, and AI is guidance, and therefore AI is good — right? 13

Not quite, and the “not quite” is the entire argument. Sweller’s own framework contains a second half that people forget: the guidance-fading effect. Worked examples help at the start. They must then be progressively withdrawn, or learning stalls — and the expertise reversal effect shows that support which helps a novice actively harms someone who has moved past that stage.

The scaffolding is only pedagogy if it comes down.

An LLM is a scaffold that never fades. It is available on the exam. It is available in the job. It is available forever, and it never asks the student to stand up without it.

And so the student never does.

The student who can call PCA and cannot derive it

Let me make this concrete, because concrete is where it bites.

A student in a machine learning course is assigned a project. They need to reduce the dimensionality of a dataset. They ask the model. It produces clean, well-commented, correct code that calls PCA(), standardises the features, plots the explained variance, and writes a paragraph interpreting the first two components. It is, honestly, better than what most students would have written. They get an A.

What did they not learn?

They did not learn that PCA is an eigendecomposition of the covariance matrix. So they do not know that it finds directions of maximum variance, which is not the same as directions of maximum information, and is definitely not the same as directions of maximum class separability — which is what they actually wanted, and which is why their downstream classifier is going to underperform for reasons they will never diagnose.

They do not know that PCA is not scale-invariant, so if someone hands them a dataset in mixed units and they forget to standardise, the component with the biggest numbers wins, and the output will look completely reasonable and be completely meaningless.

They do not know it assumes linear structure, so on a manifold it will produce confident, interpretable, wrong axes.

They do not know that the components are not identifiable up to sign or rotation, so the “interpretation” they wrote of PC1 is, in a deep sense, a story about an arbitrary basis vector.

And they do not know what to do when the covariance matrix is singular, or the data has more features than samples, or the variance is not where the signal is.

Every single one of those is a constraint. And here we meet, in a human being, exactly the failure the first essay identified in the machine: the car wash problem. The model knew that a car must be present to be washed. It failed to apply the constraint. The benchmark authors put it precisely — the failure is in constraint inference, not missing knowledge.

Knowing a fact is not knowing what the fact rules out. And the student who has never derived PCA has the fact and does not have the constraint.

They will be fine — genuinely fine — for as long as the data is conventional and the pipeline is standard and the answer is somewhere in the recorded space.

And they will be catastrophically, invisibly wrong the first time it isn’t. Which is to say: the first time it matters.

The credential dies with the signal

There is an economic consequence to all of this that universities have not begun to reckon with.

Michael Spence won a Nobel Prize for signalling theory. Education, in his model, works as a signal not because of what you learn but because the cost of acquiring the credential is lower for the able than for the unable. That cost differential is what makes the signal informative. It is what separates the types. Remove the differential and the signal carries no information — it is just paper. 18

Generative AI collapses the cost of producing the artefacts on which credentials are assessed. The essay, the problem set, the take-home exam, the semester project, the literature review, the portfolio — all of them can now be produced at near-zero cost by someone who cannot do the underlying work.

The cheat and the scholar submit indistinguishable artefacts.

Which means the credential no longer separates types. Which means — by Spence’s own logic — it stops being a signal at all.

And now watch the two halves of this essay meet. A degree becomes a lemon. The employer cannot verify it. So the employer stops paying a premium for it. So the students who would have done the hard version stop doing the hard version, because the market no longer rewards it. And the unravelling proceeds exactly as Akerlof described.

The honest steelman, and why it does not save us

Here is the strongest case against everything I have just written, and I think it is genuinely strong.

Benjamin Bloom found that one-on-one tutoring moved the average student roughly two standard deviations above classroom instruction — an effect size so large it has haunted education research for forty years, because tutoring at scale was unaffordable. 14

AI could be the best tutor ever built. Infinitely patient. Available at 3am. Able to explain the same concept eleven different ways. Able to generate an endless supply of practice problems at exactly the right difficulty. If you wanted to build a machine to deliver Bloom’s two sigma to every child on earth, it would look a great deal like this.

I believe that. I think it may be the most important positive thing this technology can do.

And it does not save us, for a reason that has nothing to do with the technology.

The same machine that can tutor the student can also do the assignment. The student chooses. And the assessment cannot tell which happened.

The tool is not the problem. The incentive is the problem. We have handed every student a device that can either give them a two-sigma education or give them an A, we have made the second option faster, easier, invisible, and free — and we have kept an assessment system that cannot distinguish the outcomes.

Nobody should be surprised by what happens next. We have not run an experiment about human character. We have run an experiment about incentive design, and we designed it badly.

The laboratory: the graduate student who does not know what was done

This is the part that frightens me most, and it is the part almost nobody is talking about, because it is happening inside institutions that are producing excellent-looking output.

The pain was the education

Before 2023, a graduate student handed a messy dataset had no choice but to live in it.

They had to open it. They had to discover that the sensor dropped out for eleven days in March, and decide what to do about it. They had to find that two of the sites used a different calibration. They had to notice that the units changed halfway through, that the missing values were coded as -999, that the timestamps were in three timezones, and that the distribution had a hard floor because the instrument couldn’t measure below it. They had to write the pipeline, watch it break, and fix it.

This was miserable. Everyone hated it. Everyone described it as the part of research that got in the way of research.

It was not getting in the way of research. It was the research.

Because at the end of those six months, that student had something no course could have given them: a physical, intuitive, almost tactile model of how that data behaves. They knew where it lied. They knew which anomalies were real and which were the instrument. They had, in the precise sense of the first essay, built the chunks — and they had built the trained perception that lets an expert see that a result is wrong before they can articulate why.

They had become the reader.

And now?

Now the student points an agent at the directory. The agent writes the pipeline. The pipeline runs. The numbers come out. The plot is beautiful. The paragraph interpreting the plot is well written.

And the student does not know what was done.

Let me be scrupulously fair here, because this is where a careless argument would overreach. In most cases, it will be fine. The agent is good. The pipeline is conventional. The data is standard. The failure modes are the ones it has seen ten thousand times. The result is correct, and the student has saved six months, and six months is a lot.

But now recall the verifier-availability law from the first essay:

AI advances where there is a cheap external verifier, and stalls where there is not.

Mathematics has Lean. Code has tests. Protein structure has the PDB.

And research — real research, the frontier — is definitionally the place where there is no verifier. That is what makes it research. If there were an oracle that could tell you whether your answer was right, you would not need to do the study.

So the situation is this:

AI is most reliable exactly where it matters least — the conventional pipeline, the standard analysis, the settled question.

And least reliable exactly where it matters most — the unconventional pipeline, the broken instrument, the model that needs reinventing, the anomaly that is either an artefact or a discovery.

And the student who let it do the easy cases has not built the capacity to handle the hard ones.

The moment the data processing fails in a way the model has not seen — the moment the pipeline is genuinely novel, the moment the neural architecture actually needs to be redesigned rather than retrieved — the person at the keyboard has no chunks, no trained perception, no critic, and no idea.

You cannot debug what you cannot derive.

And they cannot even ask

There is a failure one level worse than being unable to check the answer, and it is the one that ends research careers before they start.

They cannot ask the right question.

The first essay called this absorptive capacity, after Cohen and Levinthal: the ability to recognise the value of new information is a by-product of prior deep knowledge. A person without it cannot use information handed to them for free. 19

But it cuts in the other direction too, and this is the direction that matters in a laboratory. The good question is not a thing you think of. It is a thing you notice — and you notice it because twenty thousand hours of contact with a domain have built a model precise enough that a discrepancy feels wrong. Poincaré’s trained aesthetic sense. The chess master who sees the move. The chemist who sees that the synthesis won’t hold.

A student who has outsourced their contact with the data has no discrepancy detector. Everything looks equally plausible, because plausibility is all they have.

They will get answers. The machine is very good at answers. They will get answers to the wrong questions, forever, and they will never know.

On top of a literature already 89% wrong

And remember the ground this is happening on. The first essay noted Begley and Ellis: Amgen tried to reproduce 53 landmark preclinical cancer papers and succeeded with six. Roughly 11%. 5

That was the state of the scientific literature before we introduced unverified pipelines, executed by agents, interpreted by researchers who did not build them, at a volume no reviewer can absorb.

We are pouring an unverifiable flood into a reservoir that was already contaminated.

What happens to trust

The three failures above — market, classroom, laboratory — converge on a fourth, which is the one that finally reaches everybody.

Institutional trust is a stock. It is accumulated slowly, over generations, by institutions being right about things. And it is spent very fast.

A lawyer files a brief citing cases that do not exist, because a model invented them and he did not check. This has now happened repeatedly, in real courts, with real sanctions. A consultancy delivers a report with fabricated figures. A journal publishes a paper whose analysis nobody — not the author, not the reviewers — can actually account for. A commissioner of the European Union stands up and cites, approvingly, a study of AI’s benefits to science that had been retracted for fabrication, because nobody in the chain from the paper to the podium had the depth to catch it.

Each of these is individually survivable. What is not survivable is the second-order effect.

When a substantial fraction of professional output cannot be verified, the rational response is to stop trusting professional output.

And that response does not discriminate. It cannot. That is the entire point of Akerlof. The buyer cannot tell the good car from the lemon, so the buyer discounts every car — including the good ones. Trust collapses across the category, not selectively within it. 1

The honest expert who did the twenty years and did the work correctly is not protected by having done it correctly. They are punished for it, because they charge more and cannot prove why they should.

And inside a corporation, the version of this is quietly terrifying: an organisation whose analysts cannot verify their own analyses does not know what it knows. It has documents. It has dashboards. It has confident, well-formatted, professional-looking conclusions. It has no idea which of them are true. And it will keep operating, profitably, for years, on a foundation it has stopped being able to inspect — right up until the moment something load-bearing turns out to have been a hallucination.

Not one person in that chain will have behaved irrationally. Every individual decision — use the tool, ship the work, hit the deadline, hire the cheaper junior, don’t hire the junior at all — will have been locally correct.

That is what makes this hard. It is not a story about bad people. It is a story about a collective-action failure with no villain in it.

The incentive inversion

So far this essay has treated the counterfeit as a defensive problem — the good work getting buried, the market unravelling, the pipeline starved. But there is an active, positive force pushing in the same direction, and it may be the most powerful of all: the counterfeit is not merely tolerated. It is celebrated, funded, and held up as the model to imitate. And what a system rewards, it gets more of.

The story beats the substance, and the story is now cheap to manufacture

Capital and attention do not flow to quality. They flow to legible quality — to the thing that can be perceived as good by the person writing the cheque or the headline. And, per Akerlof, quality in expert work is often not perceptible at the moment of the decision. So the decision is made on a proxy: the demo, the deck, the narrative, the founder’s confidence, the momentum of the coverage. The proxy was always gameable. AI has made it free to game.

The pattern is visible right now, in the open. Venture rounds are priced, in the reporting of the last two years, “before there is anything to measure” — seed capital as a bet on a story, with billions raised by labs that shipped no product. First-time founders raise at scales that once belonged only to incumbents, on the strength of a thesis and a pitch. And where the story is thin, AI will thicken it: one widely discussed 2026 startup was hailed as “the fastest-growing company in history” before scrutiny revealed AI-generated marketing that included fabricated doctor profiles. The generator did not just build the product-story; it built the credentials the story leaned on.

This is the second essay’s version of the first essay’s warning. There, the worry was that AI produces artefacts a non-expert cannot distinguish from expert work. Here, the consequence: the people allocating the resources — investors, editors, hiring managers, grant officers — are, for most fields, exactly those non-experts. They are the buyers in the lemon market, and the counterfeit is now good enough, and cheap enough, to win the resources that used to require the real thing. a newcomer with no depth, a compelling narrative, and a fluent AI at their disposal can now capture the funding, the coverage, and the customers that a decade of quiet competence would once have been needed to earn.

The influencer era meets the slop machine

Two features of the present moment turn this from a problem into an accelerant.

The first is that we have moved into an attention economy governed by social proof. Legitimacy is increasingly conferred not by a field’s judgment — the slow, expert, gatekept verdict that the first essay called the source of “appreciation” — but by virality, follower counts, and the blessing of platforms whose ranking functions optimise for engagement, not truth. A field’s verdict is expensive, slow, and correct. A platform’s verdict is instant, free, and orthogonal to correctness. Guess which one now allocates attention. And AI is a virality engine: it manufactures the confident thread, the polished carousel, the plausible explainer, the outreach message, the synthetic testimonial — the whole apparatus of looking authoritative — at zero marginal cost.

The second is that the same technology has made everyone faster, busier, and more starved of the one resource verification requires: time. The promise of AI is acceleration, and the promise is being kept — but acceleration cuts both ways. As output speeds up across the board, the expectation of output rises to match it: ship faster, produce more, respond sooner. The hours available to check a thing shrink at exactly the moment the volume of things needing checking explodes. A person drowning in plausible-looking results, under pressure to move at the speed the machine sets, does not verify. They skim, they trust the surface, they pass it on. The flood does not only overwhelm the expert reader who wants to check. It removes, from everyone, the slack in which checking was possible. Deep scrutiny was always subsidised by spare time, and the spare time is being optimised away.

The vicious cycle, stated plainly

Now assemble the pieces, because they form a loop, and the loop is the real thesis of this essay.

If the counterfeit reliably captures the reward — the funding, the audience, the promotion, the contract — then the reward decouples from the depth. And a rational young person, watching where the rewards actually go, draws the obvious conclusion.

Why spend ten years becoming a real scientist, engineer, artist, or writer — a decade of deferred income and desirable difficulty and unglamorous apprenticeship — when a convincing impression of expertise, generated on demand, captures more of the reward, faster, at a fraction of the cost?

This is the incentive that governs which lives get built. When the market visibly pays more for a fluent story than for hard-won competence, fewer people will choose to acquire the competence. Not because they are lazy or corrupt, but because incentives are what a civilisation uses to tell its talented young people where to go, and the signal has been inverted. The path that builds real verifiers — long, hard, poorly paid at the start, and now visibly out-competed at the start by people using AI to skip it — becomes the irrational choice.

And here the two halves of this essay close into a single circuit:

• The people who can verify AI’s output are experts.
• Experts are produced by a long, expensive training path.
• That path was always sustained by an incentive: do the work, earn the reward.
• AI-enabled counterfeiting breaks the link between the work and the reward — the impression now earns what the substance used to.
• So fewer people enter the path.
• So the pool of people who become verifiers shrinks.
• So there are fewer experts to catch the counterfeits.
• So the counterfeits win more reliably still.
• So the incentive to enter the path falls further.

Each turn of the loop degrades the foundational pool of talent from which a society draws its scientists, its engineers, its physicians, its artists — the very people it will need to verify, correct, and direct the machine. The first essay warned that we are automating away the junior roles through which experts are trained. This is the demand-side version, and it is worse, because it operates before anyone even enters the pipeline: we are removing the reason to try.

The machine does not need to defeat our experts. It only needs to make becoming one look like a bad bet — and a system that showers funding on the confident newcomer and starves the patient apprentice is running exactly that experiment, at scale, right now.

Every completed lap, fewer of the talented choose the decade-long climb — because the reward has decoupled from the depth.

The wheel never stops on its own. Each turn makes the next turn easier — the counterfeit wins a little more, the incentive to do the real work drops a little more, and the pool of future readers falls. High price for experts, collapsing supply. That is not equilibrium.

The machine doesn't need to defeat our experts. It only needs to make becoming one look like a bad bet.

The two valuations

There is an apparent contradiction between this essay and the first one, and resolving it is not a bookkeeping exercise — the resolution is the single most important thing in the entire series, because it locates precisely where the danger lives.

The first essay ended on a triumphant-sounding fact: the price of verified human expertise is rising. The AI labs pay doctors hundreds of dollars an hour and senior strategists a thousand, to judge what their models produce. Expertise, it concluded, is the bottleneck input to AI itself, and its market value is climbing.

This essay has spent pages arguing the opposite: that expertise is being out-competed, buried, and stripped of its reward.

Both are true, and they do not conflict, because there are two different populations doing the valuing, and they are looking at two different things.

Valuation one: the people who work with the machine

The first group is the people operating closest to the technology — the labs, the serious engineering teams, the firms that have deployed these systems at scale and been burned. They have watched the model produce a confident, plausible, wrong answer and ship it into production. They have learned, at cost, the lesson of the whole series: that generation is cheap and judgment is scarce, that the output looks identical whether it is right or catastrophically wrong, and that the only thing standing between them and disaster is a human who can tell. Their valuation of expertise is expressed in money, under real stakes, with real feedback. It is a revealed preference, and it is rising, because they can see — because they have been forced to see — exactly what the expert is for. 21 22

Valuation two: the people who see only the surface

The second group is everyone else: the general public, the patient reading a chatbot’s diagnosis, the manager approving a deck, the voter, the student choosing a major, the legislator writing an education budget. This group does not work alongside the machine’s failures. They see its successes — the fluent essay, the confident answer, the polished artefact that looks exactly like expert work. And, crucially, they never see the thing the expert actually does, because the expert’s core value is the catching of errors, and a caught error is an absence. You cannot perceive the disaster that did not happen. The value of verification is, by its nature, invisible to anyone not standing at the point of verification — which is the same invisible-guardrail problem the third essay builds its warning around.

So this group draws the natural, and wrong, conclusion: the AI produced something as good as the expert, therefore the expert is replaceable. They cannot see the value of the one who evaluates and verifies, because that value never appears on the surface they are shown. Their valuation of expertise is not revealed by stakes. It is shaped by the visible artefact and the story told about it — and the story, as the previous section argued, is now cheap to manufacture and tuned for engagement rather than truth.

Why the second valuation is the one that matters

Here is the part that should command attention. The rising wage is real, but it is a lagging price in a thin, specialised market — the labs bidding for a fixed pool of existing experts. It says nothing about whether new experts are being made. And the thing that governs whether new experts are made is not the labs’ rate card. It is the second valuation — public sentiment — because public sentiment is what sets policy.

Public perception of whether expertise is worth having flows directly into the decisions that determine the supply of experts a generation from now: how education is funded, which fields a nation invests in, whether deep training is respected or dismissed as obsolete, whether a talented eighteen-year-old believes that a decade of hard study leads somewhere worth going. A society that has collectively concluded “AI replaced the experts” will defund the training, lower the status, and redirect its young — regardless of what the labs are quietly paying for verification behind closed doors. The two valuations point in opposite directions, and the one that loses the tug-of-war on price is the one that wins the tug-of-war on policy.

So the rising expert wage is not the reassurance it looks like. It is the price signal of a market discovering that verification is scarce — while, one layer up, the public and the political system that answer to it are drawing the opposite conclusion and dismantling the pipeline that produces the scarce thing. High price, collapsing supply. That is not equilibrium. That is a shortage forming in slow motion, hidden behind a wage number that looks healthy right up until the people who can command it retire and are not replaced.

And it reaches all the way to the exam room

The stakes of this are not abstract, and they are not confined to labour markets. When the public cannot distinguish the surface of expertise from its substance, the erosion reaches the most intimate and consequential encounters a person has.

A patient consults a chatbot, receives a fluent and confident answer, and concludes the doctor is unnecessary — or worse, wrong, or worse still, not worth listening to. I am not claiming the physician is always right or that the model is always wrong; the third essay is careful that a human expert is not infallible and the machine is not useless. That is not the point. The point is that a patient who does not understand what the machine cannot do — who cannot see that it produces plausible surface without the grounded, verified, accountable judgment that a trained clinician brings — will make decisions on a false model of the world, and some of those decisions will be irreversible. Multiply that across medicine, law, engineering, finance, public health, and you are no longer describing a market inefficiency. You are describing an erosion of the social basis on which trained judgment functions at all — the shared understanding that lets a society defer, appropriately, to the people who did the work.

This is why the analogy to previous technological disruption fails, and why “AI is just the next tool” is a dangerous comfort. When Amazon changed how products reach customers, it rerouted a supply chain; the goods, and the people who made them, still had to be real. What is happening now is different in kind. It is a disruption aimed not at a channel but at the mechanism by which a civilisation knows what is true — at the distinction between competence and its imitation, and at the public’s ability to tell them apart. Couple that to a flow of capital that rewards the imitation, and it is not a change in how the foundation is decorated. It is a change in the foundation. That is not a thing to embrace blindly because the demos are dazzling. It is a thing to embrace carefully, with clear eyes about exactly what these systems can and cannot do — which is the entire purpose of these three essays, and the whole of what we still control.

Training used to be bundled in. Now it is not.

I want to isolate the thing I think is genuinely, structurally new, because I have not seen it stated plainly.

For all of recorded history, professional training was a by-product of professional work.

The apprentice ground the pigments, and grinding pigments taught them about pigment. The junior lawyer read every document in discovery, and reading every document taught them what a case looks like. The graduate student cleaned the data, and cleaning the data taught them the data. The illustrator drew the storyboard, and drawing the storyboard kept the hand alive. Nobody had to pay for the training, because the training was welded to output that somebody wanted to buy.

That weld is what AI has broken. It now does the by-product task without producing the by-product.

And so training, which for centuries was free and automatic, has become a cost centre with no owner.

This is a textbook public goods problem, and it is worth naming as one, because it tells us that the market will not fix it.

• Every firm needs senior experts. They are the only people who can verify the machine.
• No individual firm has an incentive to train them, because training is now a pure cost with no output attached, and the trained expert can leave.
• The rational move for each firm is to hire experts trained by someone else.
• And so nobody trains anyone.

The seniors we are all relying on were trained by an apprenticeship system that no longer exists, funded by a market tier that no longer exists, through a struggle that has been optimised away.

We are living off inventory. And nobody is running the factory.

What would actually help

I refuse to end on doom, partly because doom is lazy and partly because I think the responses are unusually clear once the diagnosis is right. If the disease is a verification failure that produces an adverse-selection spiral, severs the training pipeline, and hides behind a public that cannot see what expertise is for, then the treatments are the ones that restore verification, restore the signal, pay for the training that used to be free, and make the invisible value of judgment visible again.

0. Make the limits of the machine common knowledge.

Before any of the institutional fixes, there is a public-understanding one, because it is upstream of all the others: the single most dangerous thing in this whole picture is a public that mistakes the surface of competence for the substance, and therefore concludes the expert is obsolete. That belief sets policy, funding, and the life choices of the next generation — and it is false, in ways that can be shown. Teach, publicly and concretely, what these systems cannot do: that they cannot reliably verify their own output, that their failures are unpredictable and uncorrelated with how confident they sound, that they are magnificent generators and unreliable judges. Not as anti-AI messaging — as basic literacy for a society now saturated with the technology. A public that understands what the machine is for and what it is not will value the human verifier correctly. A public that does not will defund the thing it most needs. The third essay is, in the end, an attempt at exactly this.

1. Bring back the viva.

If the artefact can be counterfeited, stop assessing the artefact. Assess the person, in the room, defending it.

The oral examination is not archaic. It is verification technology, and we abandoned it not because it stopped working but because it was expensive — and we could afford to abandon it only for as long as the take-home artefact remained a valid proxy. That proxy is now dead. Ask a student to derive PCA at a whiteboard, or to explain why their pipeline dropped 11 days in March, and the counterfeit collapses in about ninety seconds.

Relative to the alternative — credentials that carry no information — the oral exam is now cheap.

2. Assess the process, not the product.

Grade the derivation, the failed attempts, the debugging log, the reasoning. Not the polished output. The polished output is exactly the thing the machine produces for free, and therefore exactly the thing that no longer tells you anything.

3. Make provenance legible, and attach liability to it.

Akerlof’s own answer to the lemons problem was not despair. It was warranties, certification, licensing, and reputation — institutions that let a good seller prove they are good. The market for expert work needs the equivalent, and it needs it to have teeth: disclosure, attribution, and above all liability. Someone must be answerable for the fabricated citation. Skin in the game is not a moral nicety. It is the mechanism that makes verification happen at all — and it was one of the five properties, in the first essay, that an ensemble of AI models conspicuously lacks. 1

4. Reward the verified, not the viral.

The incentive inversion of the previous section is not a law of nature; it is a choice about what we celebrate and fund. Every institution that allocates a scarce reward — a grant, a round, a byline, a promotion, a platform ranking — is deciding whether to pay the impression of competence or the substance of it. Right now most of them pay the impression, because the impression is what is legible at the moment of the decision. The fix is to make the reward wait for the verification: fund the milestone that was actually hit, not the story about the milestone; promote the analysis that survived scrutiny, not the one that looked good in the meeting; treat “grew fastest in history” as a claim to be checked, not a headline to be printed. This is slower and more expensive than rewarding the demo, which is precisely why it stopped happening — and precisely why restoring it is the whole game. A civilisation gets more of what it rewards. If it rewards the counterfeit, it will get fewer and fewer people willing to do the real thing.

5. Pay for the apprenticeship, deliberately, because it will not happen by itself.

This is the hard one and the important one. Training used to be free. It is not free any more. If it is a public good with a free-rider problem, then it needs what public goods always need: deliberate funding, from firms, from universities, from states, from professional bodies — and it needs to be defended against the quarterly logic that will always, always cut it first, because its payoff is twenty years out and lands on somebody else’s balance sheet.

Fund the junior role that AI could do. Fund it precisely because AI could do it. That is not sentimentality. It is the cheapest insurance policy any organisation will ever buy, and the alternative is discovering in 2045 that nobody in the building can tell whether the machine is lying.

6. And, personally: choose the hard version.

Every day, in every task, there is now a version that is fast and a version that teaches you something. They produce the same artefact. Only one of them produces you.

The market will not reward you for this in year one. It will reward you enormously in year ten, because — as the first essay argued at length — when everyone’s floor rises to the same level, the floor stops being a differentiator, and the only thing left is what sits above it.

The people who did the derivation will own the next twenty years. Not because virtue is rewarded. Because they will be the only ones who can tell whether anything is true.

Back to the pile

The first essay ended in a room full of monkeys, a very good typewriter, and a growing mountain of plausible pages.

I said then that the hard part was never the typing. The hard part was always the reading.

This essay has been about how we are, right now, with the best of intentions and impeccable local logic at every step, burying the readers we have and failing to make new ones.

We answered the easy objection first — just keep an expert on hand to verify — and found it broken, because verification is not a switch, it is a scarce and often catastrophically expensive resource, and generation just went to infinity while it stayed flat. curl’s maintainers already lived it: a working verification market, functioning since 2019, shut down not because the readers got worse but because the pile got too deep to read. In the cheap-verifier domains the flood merely exhausts the experts. In the expensive ones — medicine, materials, energy — there is no cheap expert to exhaust, only a slow physical world that verifies one candidate per decade, and no amount of confident generation can hurry it.

The market that paid for their apprenticeship is being competed away by a counterfeit that the buyer cannot distinguish from the real thing — and by Akerlof’s mechanism, that does not merely undercut the real thing. It removes it from the market.

The classroom that made them struggle has handed them a machine that eliminates struggle, in a system that cannot detect the substitution, and then grades the artefact anyway.

The laboratory that made them touch the data has let them stop touching it — and the touching was the training, and the training is what would have let them notice, five years from now, that the beautiful plot is wrong.

And beneath all of it, the reason anyone submitted to that long apprenticeship in the first place is being pulled away: when the confident counterfeit captures the funding and the audience and the reward, the depth stops paying, and the talented young person watching learns that the decade of hard work is now the losing move. We are not only dismantling the machinery that makes readers. We are removing the reason to become one.

And the whole thing is happening without a villain, without a conspiracy, and without a single person doing anything that is not, from where they are standing, entirely sensible.

Here is what I think we have not understood.

We have been debating whether AI can replace experts. That was always the wrong question, and the first essay tried to show why: it cannot, because verification has no oracle outside a handful of formal domains, and where there is no oracle the verifier is a trained human being.

The right question is whether we will still be making them.

Because the machines do not need to become as good as our experts. They only need to remain in service longer than our experts do. And at the current rate — with the junior roles gone, the apprenticeships unfunded, the credentials uninformative, and the struggle optimised out of the curriculum — that is not a difficult bar.

The pile is growing. It has never grown faster. Somewhere in it, right now, there is a proof nobody has checked, a diagnosis nobody has confirmed, an analysis nobody can account for, and a fabrication that a Nobel laureate endorsed.

And the last generation of people who could tell the difference is halfway to retirement.

Predictions

Falsifiable, as before:

1. Entry-level employment in AI-exposed occupations will continue to decline relative to senior employment, and the gap will widen before it closes. (Currently: −16% for ages 22–25; +9% for 35–49.)
2. The aggregate labour numbers will look fine for years, and this will be used to dismiss the pipeline problem. It will be a mistake. A severed pipeline is invisible until the inventory runs out.
3. Assessment will move back toward proctored, oral, and in-person formats in any field that actually cares whether its graduates can do the work — and the fields that do not make this move will see their credentials lose market value first.
4. A significant institutional failure will be traced to unverified AI-produced analysis that nobody in the chain was equipped to check. (Arguably: this has already happened, at MIT and then in a European Commission speech.)
5. The wage premium for verified senior expertise will rise even as public esteem for expertise falls — the two-valuations divergence made concrete. Watch for the tell: rate cards for expert AI-evaluation climbing while surveys of public trust in professionals, and political support for funding deep training, decline in the same years. If both move together instead of apart, this prediction is wrong. (The labs’ $500–1,000/hour expert-evaluation market is the leading indicator of the first half.)
6. Firms that deliberately fund junior training through the AI transition will, in fifteen years, be conspicuously better than the ones that didn’t — and will be unable to explain why, because the counterfactual is invisible.
7. More open, unpaid verification systems will close, gate, or charge — bug bounties, open peer review, community moderation, preprint comment threads — as AI-generated volume makes free triage unaffordable. (Currently: curl’s shutdown; CycloneDX; journals piloting AI-disclosure “canaries.”) Each closure will, like curl’s, quietly raise the barrier for the next generation trying to enter.
8. A visible cohort of hype-funded, thin-substance ventures will fail publicly once the physical verifier — the market, the product actually working at scale — returns its verdict, and the gap between the funding raised on the story and the value that survived contact with reality will be large. (Early signals: the 2025 wave of well-funded AI-startup shutdowns; the “fastest-growing company in history” that turned out to rest on fabricated profiles.)
9. Enrolment and persistence in long, deep, deferred-payoff training paths — hard sciences, rigorous engineering, serious craft — will soften relative to fast-credential, high-visibility paths, tracking the perception that the reward has decoupled from the depth. This is the hardest to measure and the most important to watch, because it is the demand-side pipeline failing at the source.