The Redundant Expert: Professionals Paid to Train Their Own Replacements

For most of the industrial era, the professional was the worker technology could not reach. The factory hand could be replaced by a machine, the clerk by a spreadsheet, the switchboard operator by a network of relays. But the radiologist reading a shadow on a scan, the litigator sensing which argument would land with a particular judge, the structured-finance banker who knew instinctively that a deal was wrong before she could say why — these people were protected by something more durable than a job title. They were protected by the illegibility of their own expertise. Their judgement lived in a place no employer could copy and no algorithm could reach: inside their heads, accumulated over decades, most of it never written down because most of it could not be.

That protection is now being dismantled, and the striking thing is who is dismantling it. It is not being taken from the experts by force. It is being bought from them, by the hour, and a growing number of them are selling.

In June 2026, The Hollywood Reporter documented the entertainment-industry version of this trade: writers, editors, and development executives, squeezed out of a contracting film and television market, taking gigs through AI training platforms to correct model outputs and teach the systems to better imitate human creative judgement. That story is real, and it is wrenching, and it has been told. But it is the visible tip of something much larger and, for the professional classes, far more consequential. The same platforms that recruit out-of-work screenwriters are recruiting doctors, lawyers, bankers, and software engineers — the people whose expertise was supposed to be the last thing a machine could learn — and paying them, in some cases north of a hundred pounds an hour, to hand it over. This is not a story about creatives. It is a story about what happens to the entire idea of expertise when the cheapest way to automate a profession turns out to be to hire it.

The moat was always tacit

To understand why this moment is genuinely new, you have to understand why economists spent two decades believing it could not happen.

In 2003, the labour economist David Autor, together with Frank Levy and Richard Murnane, published what became one of the most influential papers on technology and work of its generation. Its argument, distilled, was that computers are good at tasks that can be reduced to explicit rules and bad at everything else. Routine work — the kind you can specify step by step — was exposed to automation. Non-routine work that depended on judgement, pattern recognition, and tacit understanding was safe, because you could not write down the rules for it in a form a machine could follow. The whole edifice of professional employment sat comfortably on the safe side of that line. A doctor's diagnostic intuition, a lawyer's sense of a case, an engineer's feel for a system — these were the paradigm cases of work that resisted codification.

The intellectual root of that idea is older still. In 1966 the philosopher and chemist Michael Polanyi coined the phrase that has haunted every attempt to automate skilled work since: “We know more than we can tell.” Polanyi's point was that the most valuable human knowledge is precisely the knowledge we cannot fully articulate — the surgeon's hands, the taster's palate, the reader's ear. You can watch a master at work for years and still not be able to reduce what they do to a manual. Tacit knowledge, by definition, escapes the rulebook.

For two hundred years, this was the professional's moat. Expertise commanded a premium because it was scarce, and it was scarce because it could not be transferred except slowly, expensively, and through human beings — apprenticeships, residencies, pupillages, the long grind of watching someone who already knew. You could not download a career. You had to grow one.

Reinforcement Learning from Human Feedback is, at its core, a machine for defeating Polanyi's paradox — not by finally writing down the rules, but by routing around the need to. The insight underneath RLHF is that you do not have to articulate tacit judgement in order to capture it. You only have to elicit it, over and over, at scale. Show the expert two outputs and ask which is better. Show a proposed diagnosis and ask where it goes wrong. Present a contract clause and ask the lawyer to flag the liability the model missed. The expert cannot tell you the rule, but the expert can tell you, reliably, which answer is correct — and if you gather enough of those judgements, a statistical system can infer the shape of the rule from the pattern of the choices. The tacit is made legible not by description but by demonstration, one graded example at a time.

This is the quiet conceptual violence at the heart of the story. The very quality that made expert judgement safe from automation — its resistance to being written down — turns out not to matter, because the machine never needed the rulebook. It needed the expert's verdicts. And the expert, it turns out, will supply those verdicts for a day rate.

Displacement by participation

Set this against every previous wave of technological displacement and the difference is not one of degree but of kind.

When the power loom displaced the handloom weaver, the machine did not first apprentice itself to the weaver. When the spreadsheet displaced legions of bookkeepers, it did not pay them by the hour to teach it double-entry accounting. The classic pattern of automation is substitution from the outside: a capability is built by engineers, deployed by capital, and it renders a class of workers redundant whether or not those workers participate. Their expertise is not extracted; it is simply made irrelevant. There is something almost clean about it. The weaver owed the loom nothing.

What is happening now inverts that logic. The system cannot be built from the outside, because the capability it needs — expert judgement in medicine, law, finance, screenwriting — does not exist in any codified form for the engineers to programme. It exists only inside the experts. So the only way to build the system is to bring the experts inside the process and pay them to externalise what they know. The displacement does not happen despite the worker's participation. It happens through it. The worker is not a bystander to their own obsolescence; they are, functionally, its supplier.

This is the structural novelty worth naming precisely, because it changes the moral and economic character of the whole arrangement. Call it displacement by participation. The expertise is not eliminated; it is transferred. And the transfer is voluntary in the narrow sense that no one is holding a gun to the physician moonlighting on a medical-reasoning dataset, and coerced in the broader sense that a great many of the people doing this work have arrived at it because the market for their primary expertise has already thinned beneath them, or because they can see it thinning and are hedging.

The screenwriter Ruth Fowler described exactly this hedge in a personal essay for Wired in the spring of 2026, writing about turning to AI training work because entertainment jobs had dried up and she needed cash for rent and groceries. The drying up is measurable: employment in the US motion picture and sound recording industries fell 28 per cent between July 2022 and May 2026, from 450,000 jobs to 326,000, according to Bureau of Labor Statistics figures compiled by The Guardian. Another veteran writer, quoted in the reporting on the same phenomenon, described the psychological texture of the work as something akin to a crazymaking standardised test — hour after hour of grading a machine's attempts at the thing you used to be paid to do, telling it where it went wrong, watching it get imperceptibly better on your instruction. The creative version of the story is the one that has broken through, because creative pain photographs well. But the same transaction, conducted more quietly and at higher rates, is now running through the professions that the middle class has spent a century treating as safe harbours.

The going rate for a career

The marketplace that has industrialised this transfer is best embodied by a single company. Mercor, founded in 2023 by three entrepreneurs then in their early twenties, began life as an AI-powered hiring platform and pivoted, when it discovered where the money was, into something more consequential: a broker of human expertise for the AI labs. It now connects a roster of specialists to companies including, according to legal filings, OpenAI and Anthropic, who pay for the expert feedback that trains their models. Meta was on that list until March 2026, when a supply-chain attack on the open-source LiteLLM package was used to lift some four terabytes of data from Mercor, exposing the personal details of more than 40,000 contractors along with proprietary source code and the training methodologies of its clients. Meta paused its work with the company indefinitely; OpenAI said it was investigating but kept its projects running. The contractors were not consulted about any of it, which is a fair measure of how much control the people supplying the material retain over anything drawn from them. In October 2025 the company raised $350 million in a Series C round, quintupling its valuation in eight months to roughly $10 billion. By July 2026 it was in talks again, this time at a valuation of about $20 billion — a doubling in nine months — on an annualised revenue run rate that had crossed $2 billion after doubling in four; in August, Nvidia was reported to be weighing participation. As of early September the round had not been confirmed closed.

The composition of Mercor's workforce is the part that should arrest anyone who still believes expertise is a fortress. The company has said it has more than 30,000 experts on its books, paid on average over $85 an hour, with the firm distributing more than $1.5 million to contractors every day. A proposed class action filed in a Texas federal court in May 2026, White v. Mercor.io Corp., describes that workforce with useful specificity: it includes physicians, attorneys, bankers, and software engineers. Reporting on the sector puts the rate bands into focus — physicians commanding somewhere in the region of $130 to $170 an hour on some platforms, lawyers around $110 to $130, with the highest specialist tiers in medicine, law, and finance reportedly reaching several hundred dollars an hour. The Hollywood Reporter, covering the creative end of the same market, noted writers earning up to $44 an hour and music professionals with advanced degrees up to $100.

Read those numbers as an outsider and they look generous — a good hourly wage for skilled remote work. Read them as an economist and they describe something stranger: the spot price of a profession's accumulated judgement, quoted by the hour, with no premium for the decades it took to acquire. A physician's diagnostic intuition, built through medical school, residency, and years of practice, is being purchased at a rate that does not distinguish it from the labour that produced it — because the buyer does not want the career. The buyer wants the verdicts the career produces, in sufficient volume to train them out of the physician.

Why is this the cheapest path to capability? The answer lies in the changing internal economics of building a frontier model. For most of the deep-learning era, the dominant cost was pre-training — pouring the scraped text of the internet through ever-larger networks. But the internet's supply of high-quality text is finite and largely exhausted, and the frontier has shifted to what the labs call post-training: the reinforcement and refinement that turns a fluent but directionless model into one that reasons, follows instructions, and gets specialist questions right. By industry estimates, post-training has grown from a small fraction of a frontier model's compute budget a few years ago to something like a fifth or a quarter of it today, and the human data feeding that stage has become one of the most contested inputs in the industry. Individual frontier labs are reported to be spending on the order of a billion dollars a year on human training data.

Here is the economic crux. A judgement rendered by a domain expert — the correct diagnosis, the right objection, the flawed clause caught — is worth a great deal to a model precisely because it is scarce and hard to synthesise. A single well-designed expert contribution can be worth more to a training run than a hundred cheap, generic labels, because it carries information the model cannot get anywhere else. Against the cost of the alternative — building genuine medical or legal capability by some other means, if such a means even exists — a hundred-odd pounds an hour for the real thing is not expensive. It is a bargain. The displaced expert is cheap not because their expertise is worth little, but because, at the moment of the transaction, they have very little bargaining power and the buyer has a great deal. Their profession is contracting; the platform has thirty thousand alternatives; the contract is short; the rate is take-it-or-leave-it. The expertise is priceless and the expert is disposable, and the gap between those two facts is exactly where the valuation lives — and why the number keeps doubling.

What the expert takes home

Which brings us to the question the day rate is designed to make you stop asking: beyond the money for the hours worked, what does the expert actually receive?

Consider what the transaction transfers, and what it does not return. The expert supplies the most refined product of their working life — judgement that took decades and considerable expense to acquire. That judgement is distilled into a reward model and folded into a system that can be copied infinitely, never tires, never retires, and is being marketed, in many cases, into the very market the expert came from. In exchange, the expert receives payment for the hours they sat at the keyboard, and nothing else. No equity in the model their judgement helped build. No residual when that model is licensed to a hospital network or a law firm. No attribution, because their contribution is deliberately dissolved into a statistical aggregate that bears no one's name. No ongoing claim of any kind. The relationship terminates the moment the invoice is paid, but the value they contributed does not terminate — it compounds, on the balance sheet of someone else.

This is a profoundly different bargain from the ones the more organised corners of skilled labour have fought for. When SAG-AFTRA struck in 2023, one of its central demands concerned digital likeness: the principle that if a studio wanted to scan a performer and reuse their image, the performer was owed consent and continuing compensation, not a one-time buyout. The Writers Guild won terms ensuring that AI could not be credited as a writer or used to erode writers' pay. Those frameworks are imperfect and contested, but they share a premise — that when your enduring professional essence is captured and reused, you retain a stake in it. The RLHF expert retains nothing. The structure is closer to a buyout than a licence, except that in a buyout you at least know you are selling the asset. Here the asset is your judgement, the sale is disguised as an hourly gig, and the thing being bought is your future competitive position, sold at the price of your present hour.

The conditions surrounding the work make the imbalance starker. The White v. Mercor complaint alleges that the company controls its experts' hours and weekly availability, sets their pay rates, trains and supervises them, offboards those who underperform, and monitors their work through a mandatory screen-tracking application — the indicia, the plaintiff argues, of employment rather than the independent contracting under which the workers are classified. Misclassification, if proven, means no benefits, no employer pension contribution, no protections. Layered on top are the non-disclosure agreements that the reporting describes as standard across the sector, which prevent workers from discussing their clients or, in many cases, even acknowledging what they are building. And when disputes arise, arbitration clauses frequently keep them out of open court. The worker contributes the crown jewels of their profession and receives, in return, a wage, a surveillance app, a gag, and a waiver of the right to sue in public. It is difficult to design a transaction that more efficiently separates a person from the value they create.

None of this is to say the experts are foolish for taking it. The individual calculation is often unanswerable: the rate is real, the bills are real, and the model will be trained with or without any particular physician's participation. That is precisely the logic of a collective-action problem. Each expert, acting rationally alone, has every reason to take the gig. All experts, acting together, are underwriting the erosion of the scarcity that gave their expertise its value in the first place. The tragedy is not that any one of them is making a mistake. It is that there is currently no mechanism through which they could make a different choice together.

The reasoning frontier and why expert judgement became the prize

It is worth pausing on why expert feedback specifically has become so valuable so suddenly, because it explains why the professions, and not just the arts, are now in the frame.

The generation of models that dominated the early 2020s were, for all their fluency, generalists trained to sound plausible. The competitive frontier since has moved decisively towards reasoning — models that can work through a multi-step medical differential, construct a legal argument, or debug a complex system, and get the answer verifiably right. Reasoning capability is far harder to bootstrap from scraped internet text, because the internet is full of confident wrong answers and thin on the rigorous, step-by-step expert working that distinguishes a correct chain of reasoning from a persuasive but flawed one. To train a model to reason like a specialist, you need examples of specialists reasoning correctly, graded by other specialists who can tell the difference. That is a resource the open web does not contain in sufficient quantity or quality. It exists only in the heads of practitioners.

This is why the labs have moved from crowdsourced labelling to credentialed expertise, and why the going rate for a board-certified physician's feedback has climbed so far above the rate for generic annotation. The scarcer and more consequential the judgement, the more it is worth as training signal — and the more the profession that monopolised that judgement has to lose by dispensing it. The economics reward exactly the extraction that is most damaging to the expert. A profession's tacit knowledge is simultaneously its collective inheritance and, on these platforms, its most liquid asset, sold off one graded example at a time by whichever of its members most needs the cash this month.

There is a grim elegance to the sequencing. First, a profession's public output is scraped without consent to build the base model. Then, when that model proves fluent but unreliable at the hard specialist judgements, the profession's living members are hired to supply the judgements the scraping could not capture. The first act took the profession's past; the second act rents its present; and the intended product of both is a system that competes for its future. The experts are being asked to close the one gap the machine could not close on its own — and that gap was their moat.

The problem of the last cohort

Every profession renews itself through a formation process that cannot be shortcut. A doctor is made through years of supervised practice on real patients under the eyes of senior clinicians. A lawyer is made through pupillage, clerkship, the slow accretion of cases. An engineer learns which textbook solutions fail in the field by watching them fail, alongside someone who has seen them fail before. Tacit knowledge, being tacit, can only be transmitted through immersion in the doing. This is expensive and slow, and it is the mechanism by which each generation of experts equips the next.

The experts currently selling their judgement to the training platforms were formed under those conditions. Their intuition was cultivated inside functioning professions with abundant entry-level work, mentorship, and the volume of real practice that turns knowledge into instinct. They are, in effect, a reservoir of judgement filled by an institutional apparatus that assumed the demand for that judgement would continue.

Now trace the feedback loop forward. If the models trained on this cohort's judgement succeed in performing the routine specialist work — the first-pass diagnosis, the standard contract review, the initial financial model — they will absorb precisely the entry-level and mid-tier tasks through which the next cohort of experts would have been formed. You do not become a senior radiologist without first reading ten thousand ordinary scans. You do not develop a lawyer's judgement without first grinding through the routine matters that a capable model may soon handle more cheaply. Remove the bottom rungs of the ladder and you do not merely make life harder for juniors; you sever the mechanism by which senior expertise is produced at all.

This is the deeper structural problem the professional version of the story exposes, and it is distinct from any worry about the quality of a given model's output. The judgement being harvested through expert RLHF is, in an important sense, non-renewable under the conditions the harvesting creates. The models are drawing down a stock of tacit professional knowledge that was accumulated under an apprenticeship system their own success threatens to hollow out. A profession can run for a while on its inherited stock of expertise. It cannot run indefinitely on a stock it has stopped replenishing, and the experts feeding the machine today may be the last cohort formed the old way — the last generation whose judgement was grown rather than inferred. What replaces them, if the ladder is pulled up behind them, is a genuinely open question, and none of the platforms brokering the trade have any reason to answer it.

Treating expertise as labour

If the diagnosis is that experts are being paid a day rate to transfer an asset worth vastly more than the day rate, the responses fall into a few families, none sufficient alone, all more plausible together.

The first is legal reclassification, and it has already produced its first answer. In McKinney v. Scale AI, Inc., in the San Francisco Superior Court, Scale AI agreed to pay $12.5 million to settle claims of misclassification and unpaid wages covering everyone who worked as a Contributor for Scale, Smart Ecosystem, or through HireArt while resident in California between 10 December 2020 and 28 February 2026. The settlement has only preliminary approval and, as settlements do, admits no liability, so it binds no future court. What it sets is a price. The White v. Mercor suit and a parallel action against Surge AI now put the same question — contractors, or employees in all but name? — to defendants who have watched a competitor pay eight figures rather than argue it out. If the courts find that a platform which dictates hours, monitors screens, sets rates, and offboards underperformers is running an employment relationship, the economics of the entire expert-data market shift: benefits, protections, and a higher floor follow. Reclassification will not, on its own, give the expert a stake in the model. But it would end the pretence that supplying the training signal for a hundred-billion-dollar industry is a casual side gig deserving of casual treatment.

The second family is the idea, older than the current boom, that data extracted from people is a form of labour and should be compensated and governed as such. In 2018 Jaron Lanier and the economist Glen Weyl set out the framework they called data dignity, or “data as labour,” arguing that people whose data powers digital systems are owed both compensation and collective bargaining power over its use. Their proposed vehicle was the mediator of individual data — an intermediary, something between a union and a data trust, that would negotiate on behalf of many contributors at once, because no lone individual has leverage against a platform but an organised bloc of them might. Applied to expert RLHF, the logic is direct. A physician alone cannot negotiate a residual on a medical model. Ten thousand physicians, bargaining collectively over whether and on what terms their profession's judgement may be used to train their replacements, are a different proposition entirely. The expertise that is worthless-per-hour and priceless-in-aggregate is exactly the kind of asset whose value can only be captured collectively.

That points to the third response, which is collective bargaining reaching into the gap where this work happens. The guilds and unions that won AI protections — the WGA at the studios, SAG-AFTRA on digital likeness — negotiated the relationship between workers and the employers who hire them directly. The RLHF market sits deliberately outside those relationships, in a jurisdictional vacuum where the writer training a model is not, technically, working for any studio, and the physician grading medical outputs is not, technically, practising medicine. Professional bodies — medical associations, bar associations, engineering institutions — have so far treated AI mostly as a question of tools and liability. They have not yet grasped that their members' collective judgement is now a traded commodity, or that they might have standing to govern its trade. A bar association that can discipline how a lawyer practises could, in principle, take a view on the terms under which lawyers sell their judgement to train systems that will compete with lawyers. None yet has.

The fourth response is structural and speculative: models of ownership that give contributors a durable stake rather than a one-off fee. A residual or royalty on the models an expert helped train. Equity, however small, in the value created. Data cooperatives that pool professional judgement and license it collectively on terms the members set, rather than letting a platform arbitrage each member individually. These ideas are easy to sketch and hard to build, and the platforms have every incentive to resist them, because the entire margin of the business depends on paying for the hour and keeping the compounding value. But that resistance is itself the tell. If a stake were worthless, refusing to grant it would cost nothing.

The last and least comfortable response is simply to see the trade clearly and price it accordingly. Right now the expert sells cheap partly because the transaction is disguised — framed as flexible remote work, a lifeline, a way to stay current in a shifting field, rather than as what it structurally is: the sale of a profession's future to build its replacement. An expert who understood the full nature of what they were transferring might still take the gig, because the bills are still real. But they would demand more for it, and the labs' billion-dollar data budgets suggest there is more to be had. The single most valuable thing the professions could do in the near term is refuse the framing that this is marginal work deserving of marginal terms. It is the most strategically important labour in the AI economy, and the people doing it are the only ones who can supply it.

The redundancy at the centre of the word

There is a bleak pun buried in the language here. To be made redundant, in British usage, is to lose your job because your role is no longer needed. But redundancy, in engineering, means something almost opposite: a backup, a duplicate held in reserve so the system keeps running when a part fails. The expert training their replacement occupies both meanings at once. They are being made redundant in the first sense — building the thing that will render their role unnecessary. And they are functioning as redundancy in the second sense — the human backup whose judgement is being copied precisely so that, once copied, the human is no longer required to keep the system running. The whole point of the exercise is to move the expert from the second category to the first: to extract the reserve capacity and then dispense with the reserve.

What makes this distinct from the century of automation that preceded it is that the extraction requires consent, and consent, in principle, can be withheld or conditioned. The loom did not need the weaver's cooperation. The medical model needs the physician's. That dependency is the one point of leverage the professions have, and it is closing — every graded example makes the next expert slightly more dispensable — but it has not closed yet. For this brief window, the people whose judgement the machines most need are the same people who could, if they organised, set the terms on which that judgement is transferred, or decline to transfer it on terms that offer nothing but a day rate. The economics that make expert judgement the prize also make expert refusal, or expert bargaining, unusually powerful — if it could be coordinated before the reservoir is drained.

The temptation is to end on the individual, to ask how the doctor or the writer or the banker should feel about grading a machine towards their own obsolescence. But that framing, again, lands the whole weight of a structural problem on its most precarious participants, and it lets the more comfortable parties — the labs with the billion-dollar data budgets, the platforms whose valuations keep multiplying, the professional bodies that have not noticed what is being sold from under them — off the hook entirely. The individual expert taking the gig is not the author of this arrangement. They are its raw material. The real question is not whether any given professional should sell their judgement to train their replacement. It is whether a society that depends on cultivated human expertise — in its hospitals, its courts, its firms, its studios — is content to let that expertise be quietly transferred into systems owned by a handful of companies, one hourly contract at a time, in exchange for nothing more durable than the hour. That is not a decision any single expert can make at any single laptop. It is a decision about what the professions are for, and who owns what they know. Right now, the answer being written, invoice by invoice, is that they know it only until someone pays them to hand it over.

References

1. Lesley Goldberg and colleagues, “Hollywood Workers Are Training AI Models as Job Prospects Grow Slim,” The Hollywood Reporter, 24 June 2026. https://www.hollywoodreporter.com/business/digital/ai-training-hollywood-writer-jobs-prospects-1236628302/
2. Ruth Fowler, “I Work in Hollywood. Everyone Who Used to Make TV Is Now Secretly Training AI,” Wired, May 2026. https://www.wired.com/story/i-work-in-hollywood-everyone-who-used-to-make-tv-now-training-ai/
3. “As Hollywood jobs dry up, workers are quietly training AI models to survive,” Digital Trends, June 2026. https://www.digitaltrends.com/movies/as-hollywood-jobs-dry-up-workers-are-quietly-turning-to-ai-training-to-survive/
4. Marina Temkin, “Mercor quintuples valuation to $10B with $350M Series C,” TechCrunch, 27 October 2025. https://techcrunch.com/2025/10/27/mercor-quintuples-valuation-to-10b-with-350m-series-c/
5. “AI startup Mercor now valued at $10 billion with new $350 million funding round,” CNBC, 27 October 2025. https://www.cnbc.com/2025/10/27/ai-hiring-startup-mercor-funding.html
6. “AI Startup Misclassified 30K Workers, Suit Says,” Law360, May 2026 (regarding White v. Mercor.io Corp., No. 6:26-cv-00201, N.D. Tex.). https://www.law360.com/employment-authority/articles/2475907/ai-startup-misclassified-30k-workers-suit-says
7. “AI Platform's 'Expert' Workforce Draws Misclassification Suit,” Law.com, 11 May 2026. https://www.law.com/2026/05/11/ai-platforms-expert-workforce-draws-misclassification-suit/
8. “Artificial Intelligence Firms Continue To Be Targeted for Independent Contractor Misclassification Claims,” Independent Contractor Misclassification & Compliance Blog, 9 June 2026. https://www.independentcontractorcompliance.com/2026/06/09/artificial-intelligence-firms-continue-to-be-targeted-for-independent-contractor-misclassification-claims-may-2026-ic-legal-news-update/
9. David H. Autor, Frank Levy and Richard J. Murnane, “The Skill Content of Recent Technological Change: An Empirical Exploration,” Quarterly Journal of Economics, 118(4), 2003. https://economics.mit.edu/sites/default/files/publications/skill-content-recent-technological-change.pdf
10. Michael Polanyi, The Tacit Dimension, University of Chicago Press, 1966.
11. Jaron Lanier and E. Glen Weyl, “A Blueprint for a Better Digital Society,” Harvard Business Review, 26 September 2018. https://hbr.org/2018/09/a-blueprint-for-a-better-digital-society
12. “What is data dignity?” TechTarget, definition and explainer. https://www.techtarget.com/searchenterpriseai/definition/data-dignity
13. “AI and Job Postings: From Destruction to Creation?” Indeed Hiring Lab, 8 July 2026. https://www.hiringlab.org/2026/07/08/ai-and-job-postings-from-destruction-to-creation/
14. “RLHF Explained: How Human Feedback Trains AI Models in 2026,” Decode the Future. https://decodethefuture.org/en/rlhf-explained/
15. “Data Annotation for AI Labs: Recruiting Guide 2026,” HeroHunt. https://www.herohunt.ai/blog/data-annotation-ai-labs-the-recruiting-guide-2026/
16. “30,000 Professionals Are Training Their Own Replacements,” Metaintro, 2026. https://www.metaintro.com/blog/mercor-ai-training-white-collar-jobs-skills-2026
17. “Mercor: Unlocking Human Potential in the AI Economy” (Series C announcement), Mercor, 27 October 2025. https://www.mercor.com/blog/series-c/
18. ”'Digging the grave of my profession': the Hollywood creatives training AI to do their jobs,” The Guardian, 22 August 2026. https://www.theguardian.com/technology/2026/aug/22/the-hollywood-creatives-training-ai-to-do-their-jobs
19. Marina Temkin, “Mercor is in talks for a $20B valuation,” TechCrunch, 9 July 2026. https://techcrunch.com/2026/07/09/mercor-is-in-talks-for-a-20b-valuation/
20. “Nvidia Weighs Investment in Round Valuing Mercor at $20 Billion,” PYMNTS, 19 August 2026. https://www.pymnts.com/news/investment-tracker/2026/nvidia-weighs-investment-in-round-valuing-mercor-at-20-billion/
21. “After data breach, $10B-valued startup Mercor is having a month,” TechCrunch, 9 April 2026. https://techcrunch.com/2026/04/09/after-data-breach-10b-valued-startup-mercor-is-having-a-month
22. “Scale AI Misclassification Settlement: California Class Action” (McKinney v. Scale AI, Inc., S.F. Super. Ct. No. CGC-24-620481, preliminary approval 2026), King & Siegel LLP. https://www.kingsiegel.com/blog/scale-ai-misclassification-settlement-california-class-action/

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...