Four Million Dollars a Day: The Experts Teaching AI to Replace Them

August 6, 2026

There is a moment, described more than once by the people who have lived it, when the correction stops being needed. You are a speech and language pathologist, say, and you have spent a fortnight teaching a language model how a sentence should breathe in Brazilian Portuguese, how a piece of creative writing earns its emotional turn rather than announcing it. You mark the model's output. You explain, patiently, why the phrasing is wrong, why a native ear would flinch. And then one morning you sit down to the same task and find you have nothing to say. The model has absorbed you. The thing you were correcting a week ago it now does on its own, in your voice, without you. Carolina Perez Sands did exactly this work for the artificial-intelligence training company Mercor, and she described the arc of it to The Wall Street Journal's podcast in June 2026. Her corrections, she said, became unnecessary within roughly a week. She has since left the industry, having reached a conclusion that ought to be printed on the wall of every laboratory in the sector. Her job, she decided, was “actually making this more of a monster.”

That sentence is the whole subject of this essay, and it is worth sitting with before we reach for the economics. She did not say the work was hard, though it was. She did not say the pay was poor, though for many in the sector it becomes so. She said the work was monstrous, and she meant something precise by it: that the better she did her job, the less of her there was left to do. She was not being replaced by a machine in the ordinary sense, the sense in which a loom replaces a weaver or a spreadsheet replaces a clerk. She was pouring herself into the machine that would replace her, decanting a career's worth of tacit judgement into a system engineered to make that judgement free and infinite. And she was being paid, by the hour, to do it.

The supply chain of the mind

The company that paid her is three years old. In July 2026 The New York Times, in a report by Lora Kelley bluntly titled “The Work of Helping A.I. Destroy Work,” laid out the scale of the operation with a single arresting figure: every day, Mercor pays more than thirty thousand contractors upward of four million dollars to help make their own jobs, and the jobs of their colleagues, obsolete. This is not the old data economy of blurred street signs and flagged slurs. The postings the Times examined read like a fever dream of the professional class: two hundred and twenty-five dollars an hour for a voice actor who could hold a customer-service persona in fluent Hebrew, a doctorate-holding physicist with a specialism in general relativity, astrophysics or cosmology, physicians who could describe the texture of primary care in Rwanda. Mercor supplies mathematicians to annotate formal proofs, lawyers to mark up briefs, professors to grade essays. Data labelling, as the Times put it, has moved up the value chain.

The money follows. Mercor was valued at ten billion dollars in October 2025 after a funding round of some three hundred and fifty million, an event that made its three founders — Brendan Foody, Adarsh Hiremath and Surya Midha, Thiel Fellows all, and none of them much past twenty-two — among the youngest self-made billionaires alive. By July 2026 Bloomberg, Forbes and TechCrunch were reporting the company in talks to raise a further five hundred million dollars or so at roughly double that valuation, twenty billion, on the back of a gross revenue run rate that had itself doubled, in the space of about four months, to reach two billion dollars a year. The doubling is the fact worth holding on to: the valuation doubling in nine months, the revenue in four. The founder Brendan Foody has offered his own version of the daily wage bill, putting it at over one and a half million dollars a day; the Times, counting more broadly, put it above four million. Either way the arithmetic is the same in shape. A very small number of very young people have built an extraordinarily valuable company whose principal activity is buying, by the hour, the accumulated expertise of tens of thousands of highly credentialled professionals, and selling it onward to the laboratories building the models that will render that expertise abundant.

It is worth being clear about how new this is, because the novelty is the whole argument. The AI industry has always rested on hidden human labour. For most of the last decade that labour was cheap, distant and largely invisible: the workers in Nairobi and Gulu paid a dollar or two an hour by outfits contracting for Scale AI and OpenAI to tag images, moderate horrors and rank chatbot replies. Nearly a hundred of them — ninety-seven, to be exact — wrote to the American president in May 2024 describing their conditions as amounting to modern-day slavery; researchers documented the psychological toll of the content they were made to sift; Scale AI, according to reporting from the region, moved to disband labeller organising in Kenya in 2024. That economy has not vanished. But Mercor represents its mirror image and its escalation at once. The people being paid now are not the world's poorest but among its most expensively trained, and what is being extracted from them is not attention or a strong stomach but the very thing their years of study were supposed to make scarce and valuable: judgement.

What the model is actually buying

To understand why this matters more than the raw injustice of any single wage cut, you have to understand what is being sold, and it is not what the word “data” suggests. When a mathematician annotates a proof for a model, she is not handing over a fact that could be looked up. She is handing over the shape of her reasoning: which steps are load-bearing and which are decoration, where a student would go wrong, what an elegant move looks like as against a merely correct one. This is what the philosopher Michael Polanyi called tacit knowledge, the knowledge captured in his famous formula that we know more than we can tell. It is the knowing-how that lives beneath the knowing-that, the reason an expert can recognise in a glance what she could not fully explain in an afternoon. It is, precisely, the part of expertise that cannot be written in a textbook, which is why professions have always transmitted it through apprenticeship, through years of supervised proximity to someone who already has it.

The entire proposition of the expert-annotation economy is that this tacit layer can, in fact, be told — extracted, structured, and used to train a system that reproduces it. The physician describing Rwandan primary care is not uploading a medical fact. She is externalising the clinical intuition she built over a decade of patients, the pattern-recognition that lets her weight a symptom differently depending on a context no guideline captures. The lawyer marking up a brief is teaching the model where the argument is weak in a way only a practitioner would feel. Each correction is a small act of translation, turning the untellable into the tellable, and the astonishing, disquieting discovery of the last two years is how few such translations the models now need before they can do without the translator. Perez Sands measured hers in a week.

This is where the MIT Sloan management professor Danielle Li put her finger on something more fundamental than any one person's redundancy. Writing in the Financial Times in March 2026, Li argued that the threat here is not merely to jobs but to the deep structure of economic security itself. “Historically,” she wrote, “economic security has rested on the scarcity of skill.” That is the sentence to underline. The reason a radiologist or a litigator or a structural engineer could command a good wage and a stable life was not only that their skill was valuable but that it was rare, slow to acquire, and lodged in scarce human heads. Once a top performer's judgement is codified into a model, Li observed, it can be copied out to every other worker in that role, everywhere, faster than any chain of human mentorship could ever manage. The scarcity that underwrote the wage evaporates. Skill does not become worthless; it becomes ubiquitous, which for the person who used to be paid for its rarity amounts to much the same thing.

The arc every worker describes

There is a grim regularity to the testimony coming out of this sector, a shared narrative shape that recurs across specialisms and platforms with the fidelity of a natural law. It begins with the pay, which looks, at first, wonderful. Moneywise reported white-collar contractors being drawn in at two hundred dollars an hour to train models on their own professions, sums that dwarf what many of them earn in their day jobs. Then the timers appear. Tasks that were open-ended acquire deadlines; the deadlines tighten; the rate per task, quietly, falls. The feedback, once collegial, curdles into something vague and demoralising, a stream of rejections whose logic is never quite explained. And then, within months, the professional notices the thing that ends the story: the model has got good. As Amanda Brown, an assistant professor of biology at Tarleton State University, told the Times, she began to notice improvements so rapid that it grew “trickier to find things the AI models didn't already know” — which is to say her own knowledge was being exhausted, mined out, the seam running thin. The work intensifies precisely as it becomes futile.

The economics of this squeeze are not accidental, and one need not impute malice to see the mechanism at work. An arXiv paper published in April 2026 by Ana-Andreea Stoica, Celestine Mendler-Dünner and Moritz Hardt, working between the Max Planck Institute for Intelligent Systems and the Tübingen AI Center, describes the general logic with unsettling clarity. A platform that controls task allocation can exploit workers' uncertainty about the true cost of their own labour to drive the effective wage down, and can wait out any collective resistance simply by reassigning the task to whoever will accept the lowest price first. What the authors prove is starker than the intuition suggests: a platform can get every one of its tasks completed while paying only a vanishing share of what that labour is actually worth — a share that shrinks as the pool of available workers grows, falling away in proportion to the logarithm of the pool divided by its size. Add workers, and the fraction of the true cost the buyer must pay dwindles towards nothing. Solidarity, on this account, requires that everyone hold the line; the platform needs only one person to break it, and with a global pool of the credentialled and the anxious, there is always someone who will. The worker's leverage — the thing a union exists to pool and protect — is dissolved by an architecture that lets the buyer transact with the single most desperate seller at any moment.

But the paper's title carries a second clause, and it is the more important one: stochastic wage suppression on gig platforms, and how to organise against it. Having shown how thoroughly the mechanism works, the same authors ask what defeats it, and the answer they find is neither utopian nor expensive. A coalition of workers committed to a price floor can force the platform's total outlay from that vanishing logarithmic share up to a linear one — to something proportional, that is, to the real cost of the work — but only if the coalition is chosen with precision. Organise a small, targeted group of the lowest-cost workers, the very people the platform is relying on to undercut everyone else, and its ability to wait out the line collapses. Draw a group of the same size at random from the same workforce and almost nothing happens: the platform routes around them and the wage keeps falling. That asymmetry inverts the folk wisdom of the sector. Resistance to an algorithm holding all the cards is not futile; undifferentiated solidarity is. Numbers alone are not power. Position is. A movement that recruits broadly among the comfortable and the visible but never reaches the bottom of the price distribution will simply be bypassed, while one that begins at the bottom, where the platform's leverage actually lives, does not have to be large to bite.

Mercor has already furnished a concrete illustration of the harder edge of this. In November 2025 Forbes reported that the company abruptly cancelled a project called Musen, on which thousands of contractors — more than five thousand of them at its peak — had been reviewing content, and then offered to rehire them at sixteen dollars an hour rather than the twenty-one they had been getting: a cut of nearly a quarter, and a rate below the legal minimum wage in California, Washington and Connecticut. Contractors described being locked out of their Slack, told the work they had been promised through December was over, and invited back at the lower price the same afternoon. The replacement project had a name, Nova, and, by the account of contractors who worked on both, tasks near enough identical to the ones they had been doing the week before, for five dollars an hour less. The company's stated rationale, delivered by email, was that the new rate would offer them “greater earning stability and consistent access to work.” “It felt like a slap in the face,” one contractor said. “We are working with AI but we don't work for AI.” The episode is a small, sharp emblem of the whole arrangement's asymmetry: the value the workers created flowed upward into a ten-billion-dollar valuation, while the risk and the caprice flowed down onto them.

Why this is not simply gig work

It is tempting to file all of this under the familiar heading of precarious platform labour, alongside the food couriers and the ride-share drivers, and to reach for the familiar remedies: minimum rates, misclassification suits, the slow grind toward employee status. Those remedies matter, and the wave of class actions already filed against Mercor in the wake of a March 2026 data breach suggests the ordinary machinery of labour law is beginning, belatedly, to turn. But to stop there is to miss what makes knowledge-extraction labour a genuinely distinct category, and the distinction is not sentimental. It is structural, and it turns on the difference between selling your time and selling your replacement.

It is worth pausing on what those suits allege, because the detail bears directly on the argument. On or about the twenty-fourth of March 2026, a threat group known as TeamPCP exploited a vulnerability in LiteLLM, an open-source library that thousands of companies use to connect their applications to commercial AI models — a supply-chain attack that caught Mercor along with much of the rest of the industry. Roughly four terabytes of data left the company: some two hundred and eleven gigabytes of candidate records, including CVs, verified contact details and Social Security numbers; around three terabytes of video and identity-verification material, including recorded interviews and images of government identification documents; and a further nine hundred and thirty-nine gigabytes of source code and internal systems data. At least seven class actions have since been filed in federal courts in California and Texas. And among their allegations is one that ought to be read alongside everything else in this essay: that Mercor secretly surveilled its contractors using screenshot-capturing software, so that what escaped included not only the documents the workers had submitted but images of their screens while they worked. The company disputes the claims and says it complies with all applicable regulations.

That is the algorithmic management this essay has been describing, rendered literal. The contractors were not merely priced by an algorithm and reassigned by one; they were watched by one, at intervals they did not control, in a stream of images they never saw and could not audit, which then passed into the hands of strangers. Meta paused its work with Mercor indefinitely. OpenAI opened a review and, along with Anthropic, stayed. And three months after a breach that cost the company one of its largest clients, it was in talks to double its valuation to twenty billion dollars. That sequence is the essay's argument in miniature. The workers whose government identification and Social Security numbers were exposed bore the risk of an arrangement they did not design; the valuation doubled regardless. Risk flows downward, and consequence, when it lands at all, lands somewhere other than where the decisions were made.

Return, then, to the structural distinction, because it is the thing the ordinary remedies cannot reach. When a courier delivers a meal, the value of that labour is consumed in the moment and gone. The platform is no closer to being able to deliver the next meal without a courier; tomorrow it must hire one again. The relationship, exploitative as it may be, is at least recurring, and recurrence is the ground on which all worker power has ever been built. The threat to withdraw labour has teeth only because the labour is needed again. But the annotation of expertise is not consumed in the moment. It is captured, retained, and compounded. Each correction Perez Sands supplied made the next correction less necessary, until the whole category of her labour was no longer needed from anyone. This is not a job; it is a liquidation. The worker is not renting out her skill by the hour. She is selling the freehold, in instalments small enough that she may not notice she has signed away the deed until the last one clears.

That difference has a further consequence that ordinary gig work does not carry. A profession is not merely a stock of individuals; it is a system for reproducing itself, generation after generation, through the apprenticeship of the young by the experienced. Law, medicine and engineering have always insisted that their craft cannot be learned from a book, that it must be absorbed through years of supervised drudgery on real problems. The expert-annotation economy attacks this pipeline from both ends at once. It codifies the judgement of today's masters into models that make tomorrow's apprentices look redundant before they are hired, and in doing so it removes the junior tasks through which mastery was ever acquired. A profession that stops being able to pay its novices stops being able to make its experts, and a machine trained on the last generation of experts has no obligation, and no ability, to grow the next. What is being extracted, then, is not only the individual's future income. It is the profession's capacity to exist.

The uncomfortable case for abundance

Honesty requires a hard turn here, because there is a powerful argument on the other side, and an essay that suppressed it would be propaganda. The scarcity of skill that Danielle Li identifies as the historical basis of economic security is, from another angle, simply a shortage — and shortages are not obviously things a decent society should want to protect. That radiological expertise is rare is wonderful for radiologists and terrible for everyone in a country that does not have enough of them. If the tacit judgement of an excellent physician can be genuinely codified and distributed, then a child in a rural clinic with no specialist for two hundred miles might receive something approaching expert care. The Rwandan primary-care knowledge being annotated into a model could, in principle, be delivered back to Rwandan clinics that have never had enough doctors. To be against the diffusion of expertise as such is to be against the thing that medicine, education and law claim, in their better moments, to want: their own universality.

There is a second concession to be made here, more concrete than the first, and it should be stated plainly rather than buried in a subordinate clause. Mercor is not, on the available figures, a company skimming most of the value off its workers' labour. The two-billion-dollar run rate is gross billings, and contractors are reported to take home somewhere between sixty and seventy per cent of what the company charges its clients, leaving Mercor's own net revenue nearer six or eight hundred million. Judged as a middleman's cut, that is a generous one — more generous than a good many staffing agencies, literary agents and record labels manage — and any argument that proceeds as though the workers were being fleeced on the split is arguing with a company that does not exist.

This is the genuine moral complication, and it cannot be waved away by pointing at the founders' valuations. Nor does the revenue share answer the objection, generous though it is. Sixty or seventy per cent of an hourly rate is still payment for an hour. It is a share of the wage bill, not a share of the asset, and the percentage cannot reach the question the arrangement actually raises: what the worker is owed for the durable thing that survives the hour, the codified judgement that goes on earning long after she has stopped, and what becomes of her when the project ends, as Musen ended, and the model no longer needs the correction she was hired to supply. A fair price for time is not a wrong thing. It is simply an answer to a different question. The problem with the expert-annotation economy is not that it diffuses skill. Diffusing skill is, at least potentially, a public good of the first order. The problem is who captures the value released by that diffusion, and on what terms the people whose lives are dismantled in the process are treated. There is nothing in the technology that requires the surplus from making expertise abundant to flow, untaxed and unshared, to three men in their early twenties and their investors, while the professionals who supplied the expertise are managed by algorithm into ever-lower rates and then discarded when their knowledge is spent. The abundance and the injustice are separable. The industry's rhetorical trick — and it is the same trick performed by every disruptive technology before it — is to bundle them, so that any objection to the injustice can be dismissed as an objection to the abundance, a Luddite's fear of progress. It is not. One can want the child in the rural clinic to have the model and still insist that the doctor who trained it was owed something more than a tightening timer and a wage cut below the legal floor.

The right question, then, is not whether expertise should be diffused, but whether the current mechanism of diffusion is the only one available, or merely the one that happens to concentrate the gains most efficiently at the top. Framed that way, the appeals to inevitability lose their force. Nothing about training a general-relativity model requires that the physicist be paid piece-rate through an app that can reassign her task to the lowest bidder mid-project. That is a choice about the distribution of power and reward, dressed as a fact of nature.

There is a legal asymmetry buried in all this that sharpens the injustice further. When the mathematician annotates a proof or the lawyer marks up a brief, the resulting model weights become the intellectual property of the company that commissioned the work, protected, very often, as a trade secret, an asset the firm can guard, license and sell in perpetuity. The expertise that went into it enjoys no such protection running the other way. The professional retains no residual claim, no royalty, no acknowledgement in the artefact her judgement helped to build; the value she supplied is enclosed the instant it is captured, converted from her tacit possession into someone else's proprietary estate. The law is thus mobilised on one side of the transaction and silent on the other. It recognises the model as property worth defending while treating the human judgement distilled into it as a spent input, like electricity or compute, with no continuing interest in what it becomes. That imbalance is not a natural feature of knowledge. It is an artefact of which parties had lawyers when the terms were written, and it could be written differently.

What is owed, and to whom

Begin with what is owed to the workers, because it is the most tractable. The economist and technologist Jaron Lanier, with the political economist Glen Weyl, proposed in a 2018 Harvard Business Review essay a framework they called data dignity, or data as labour. Their insight, made years before the present moment but fitting it exactly, was that the digital economy systematically mislabels as free capital what is in fact human labour — the data and judgement that people supply to the systems that profit from them. Their remedy was not to ban the practice but to price it honestly: to treat the supply of training value as work, to be compensated as work, and crucially to be bargained over collectively. They imagined intermediary bodies — mediators of individual data — that could negotiate royalties and terms on behalf of the people supplying the value, much as a guild or a union once did. Danielle Li reached, from the other direction, a strikingly similar practical conclusion: if your work is training a model that will benefit your employer or a buyer, you should seek explicit recognition and payment for that, rather than assuming your ordinary wage already covers the sale of your professional soul.

The mechanism that makes this urgent rather than merely fair is the information asymmetry the Max Planck researchers described. A worker who does not know that her fortnight of corrections will make her whole role redundant cannot price that fortnight correctly. She is selling an asset — the future scarcity of her skill — without being told that is what is on the table, and at a price set by a party who knows exactly what it is worth. This is the classic condition under which markets fail and regulation earns its keep. At minimum, knowledge-extraction contracts should carry something like informed consent: a disclosure that the labour being purchased is training a system intended to perform the worker's function, and terms — royalties, residuals, equity, a share of the model's downstream value — that reflect the durable nature of what is being transferred rather than treating it as spent the moment the hour ends. Actors and writers won residual rights and consent provisions over their digital likenesses through collective action; there is no principle that grants a screen actor a stake in her synthetic double but denies a physicist one in the model built from her mind.

None of this is hypothetical, and the working proof of it comes from the other end of the supply chain rather than the top. In Kenya, the labellers whose two-dollar-an-hour work built the previous generation of these systems formed the Data Labelers Association, which signed up three hundred and thirty-nine members in its first week and has been pressing since for a code of conduct binding on the major labelling platforms: equitable pay, freedom of association, scheduled breaks, psychological support for the people made to sift the worst material on the internet. It has weighed legal action against Remotasks over the sudden withdrawal of platform access and wages left unpaid. It is small and under-resourced, and it is very nearly the body Lanier and Weyl imagined, assembled without their help at the poorest end of the chain, where the leverage is thinnest and the risk of organising highest.

The Max Planck result explains why that may be the right place to start rather than merely the most desperate. If a precisely targeted coalition of the lowest-cost workers is the thing that forces a platform to pay something approaching the real cost of labour, while a coalition of the same size drawn at random achieves almost nothing, then organising that begins in Nairobi is not a sideshow to whatever a displaced radiologist or general-relativity physicist may one day contemplate. It is the load-bearing part. The credentialled professional in California and the labeller in Kenya are not two separate stories about AI and work; they are the top and the bottom of a single price distribution, and it is the bottom that determines what the top can hold out for. Solidarity across that distance is not sentiment, and it is not charity. On the mathematics, it is the only version that works.

There is a second, blunter instrument that belongs in the same toolbox: the pooling of the transition's risk rather than its wholesale offloading onto the individual. If the diffusion of expertise generates a genuine productivity windfall — and the valuations suggest the market believes it does — then some of that windfall can be recycled into the people it displaces, through wage insurance that tops up the earnings of a professional who must move to lower-paid work, through funded retraining that is more than a gesture, through direct support during the months when a codified skill is losing its market. None of this is exotic; versions of it already exist for workers displaced by trade. The point is that the surplus released by making a profession abundant need not vanish, untaxed and unshared, into a balance sheet. It can be treated as the collective product it partly is, and distributed accordingly. What turns displacement from a catastrophe the worker absorbs alone into a risk the society that benefits agrees to pool is nothing more mysterious than the decision to share.

What is owed to the professions is subtler and harder. A profession is a public trust as much as a private career; society licenses doctors and lawyers not merely to protect their incomes but to guarantee that the expertise exists at all, renewably, accountably, with someone who can be struck off. When the expert-annotation economy hollows out the apprenticeship pipeline, it privatises a capacity that was always partly public, transferring the reproduction of medical or legal judgement from a regulated profession to an unregulated model owned by a private company. The obligation here runs to the institutions that credential and govern professions: to insist that the tacit knowledge being harvested from their members is not simply enclosed, that models trained on a profession's collective judgement carry some duty of stewardship back to it, and that the training of the machine does not quietly defund the training of the humans on whom the machine will always, ultimately, depend for correction, contest and renewal.

Governing the extraction differently

So, to the question the commissioning editor poses directly: should the labour of knowledge extraction be governed differently from other forms of gig work? The answer this essay reaches is a qualified yes, and the qualification is as important as the assent, because the case for special treatment rests not on the workers' credentials but on two structural features that ordinary gig work does not share.

The first is the terminal, self-cannibalising nature of the labour. Most work is repeatable; this work is designed to eliminate its own recurrence, and labour that abolishes itself cannot be protected by the standard remedies, which all assume a continuing relationship in which power can be exercised. You cannot strike a job that will not exist next month. This is why minimum-rate rules and misclassification suits, necessary as they are, cannot be the whole answer: they regulate the terms of a relationship whose defining feature is that it is engineered to end. Governing this labour honestly means attaching value to what is captured rather than only to the hours spent capturing it — residuals, collective royalties, a claim on the diffused asset — precisely because the hourly frame is the mechanism of the dispossession.

The second is the acute information asymmetry, sharper here than in any courier's contract, over what is actually being sold. When the buyer knows and the seller does not that the transaction extinguishes the seller's future market, the ordinary presumption that a freely struck bargain is a fair one collapses. That is a textbook justification for mandated disclosure and for a floor of non-waivable rights, the same logic that governs financial advice and the sale of securities. A society that requires a mortgage broker to disclose a conflict of interest can require an AI-training platform to disclose that the task on offer is the codification of the worker's own obsolescence.

What it does not justify is protectionism dressed as principle. The temptation, for a displaced professional class newly acquainted with the underside of technological change, will be to defend the scarcity of skill for its own sake — to treat the diffusion of expertise as a harm to be prevented rather than a good to be shared. That way lies a defence of shortage, and shortage is not justice; it is merely the old distribution of luck. The obligation is not to keep expertise rare so that its holders can keep charging rents. It is to ensure that when expertise is made abundant, the people from whom it was taken are treated as the authors of a public good rather than the raw material of a private one — compensated durably, informed honestly, bargaining collectively, and not managed by algorithm into pricing their own erasure at a discount.

The monster and the mirror

Return, at the end, to Carolina Perez Sands and the word she chose. She did not call the work exploitation, though a wage below the Californian minimum would qualify. She called it monstrous, and the horror she named was not primarily about money. It was about complicity — the vertiginous recognition that her own excellence was the instrument of her erasure, that every good correction she made was a brick in the wall being built between her profession and its future. She left. Most cannot, and the platform is designed on the assumption that for every one who walks away in disgust there is another, somewhere in the global pool of the credentialled and the underemployed, who will take the task at a lower price and never know how little of themselves they are selling until it is gone.

The economy she describes is not a marginal curiosity. It is, on the evidence of Mercor's valuation, one of the fastest-growing businesses in Silicon Valley, and it represents in concentrated form the central bargain of the AI transition: the conversion of scarce, hard-won, human judgement into abundant, ownable, machine capacity, with the gains flowing to whoever owns the machine and the losses absorbed by whoever supplied the judgement. There is a version of that conversion that is a gift to humanity, the expert diffused to every clinic and classroom that never had one. And there is the version we are building, in which the diffusion is real but the dividend is captured, and the experts are paid by the hour to dig their own seam until it is empty. The technology does not choose between these. We do, through the terms we set, the disclosures we require, the bargaining we permit, and the share of the surplus we insist flows back to the people whose minds made it possible.

Danielle Li was right that economic security has always rested on the scarcity of skill, and right, too, that this foundation is dissolving. But the correct response to a dissolving foundation is not to mourn the scarcity. It is to build a new basis for security that does not depend on shortage — one in which the diffusion of what we know enriches the people who knew it first rather than discarding them. That is a political choice, not a technical one, and it is still, for a little while longer, ours to make. The monster is not the model. The monster is the arrangement whereby we pay the wisest among us, by the hour and below the wage floor, to feed themselves to it, and call the result progress. We can build the abundance without the monster. We are simply, at present, choosing not to.

References & Sources

Lora Kelley, “The Work of Helping A.I. Destroy Work,” The New York Times, 10 July 2026.
“AI Training Startup Mercor Discusses $20 Billion Valuation,” Bloomberg, 9 July 2026.
“Mercor is in talks for a $20B valuation,” TechCrunch, 9 July 2026.
Richard Nieva, “AI Data Labeler Mercor In Talks To Raise $500 Million At $20 Billion Valuation,” Forbes, 9 July 2026.
“Mercor doubles to $2B gross revenue run rate as AI labs buy expert data,” Dealroom, July 2026.
Iain Martin, “The World's Youngest Self-Made Billionaires Just Slashed These Workers' Wages By A Third,” Forbes, 12 November 2025.
Hugh Langley, Grace Kay and Shubhangi Goel, “An AI startup powering Meta and OpenAI cut thousands of workers — then offered them a similar project for less money,” Business Insider, 12 November 2025.
“The Journal” podcast, The Wall Street Journal, interview with Carolina Perez Sands, June 2026.
Danielle Li, opinion essay on artificial intelligence and economic security, Financial Times, March 2026.
Ana-Andreea Stoica, Celestine Mendler-Dünner and Moritz Hardt, “Stochastic wage suppression on gig platforms and how to organize against it,” Max Planck Institute for Intelligent Systems, Tübingen AI Center and ELLIS Institute Tübingen, arXiv:2604.15962, 17 April 2026; published in the Proceedings of the ACM Web Conference 2026.
“Mercor pays over $1.5 million a day to humans training AI, says its CEO,” Yahoo Finance / Reuters, 2026.
“White-collar workers are getting paid $200 an hour to train AI on their jobs — but they say it's not 'easy money',” Moneywise, 2026.
Jaron Lanier and E. Glen Weyl, “A Blueprint for a Better Digital Society,” Harvard Business Review, September 2018.
“AI is a multi-billion dollar industry. It's underpinned by an invisible and exploited workforce,” The Conversation, 2024.
Open letter from data labellers, content moderators and AI workers in Nairobi to President Joseph R. Biden, 22 May 2024.
“Kenyan AI workers form Data Labelers Association,” Computer Weekly, February 2025.
“Mercor says it was hit by cyberattack tied to compromise of open source LiteLLM project,” TechCrunch, 31 March 2026.
“Mercor, a $10 billion AI startup, confirms it was the victim of a major cybersecurity breach,” Fortune, 2 April 2026.
“AI staffing firm Mercor faces lawsuits over data breach,” Staffing Industry Analysts, 2026.
“Meta pauses work with AI data firm after security incident,” Computing, 2026.
Michael Polanyi, The Tacit Dimension (Routledge & Kegan Paul, 1966).
“Mercor Mission — Organizing human intelligence to power the AI economy,” Mercor, 2026.

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Discuss...