SmarterArticles

There is a particular kind of vertigo that comes from being asked to teach a machine the thing that used to be your livelihood. You sit at a laptop, often under a nondisclosure agreement, often managed by someone two decades your junior, and you read what an artificial intelligence has written. Then you tell it where it went wrong. You score the dialogue. You flag the moment the scene lost its rhythm, the joke that did not land, the emotional beat that rang false. You are very good at this, because you spent a career learning to do exactly this for human writers, human editors, human rooms. And the better you do it, the more you sharpen a system that already stands accused of hollowing out the market you came from.

This is not a hypothetical. According to a 24 June 2026 investigation by The Hollywood Reporter, it has quietly become a survival strategy for a meaningful slice of the American entertainment workforce: television writers, film editors, and development executives who can no longer find enough traditional work are taking gigs training the very generative-AI models widely understood to be contributing to the contraction of that same market. They are, in the publication's framing, helping to “smooth out” the systems that may pave the way to their own obsolescence.

The most arresting line in the reporting came from a veteran writer who had spent time inside this work. “How many times can you tell a machine it's wrong without losing your mind?” The question is not rhetorical. It is, increasingly, an occupational hazard.

The shape of the bargain

To understand why a mid-career screenwriter would take this work, you have to understand that for the person who needs to pay rent this month and cannot find traditional employment, the decision is not philosophical. It is financial. That distinction matters enormously, and it is the first thing that gets lost when the arrangement is described from the outside as a betrayal.

Writer Ruth Fowler put the calculus plainly. In a personal essay published by Wired in May 2026 — bluntly titled “I Work in Hollywood. Everyone Who Used to Make TV Is Now Secretly Training AI” — she described turning to AI training because entertainment work had dried up and she “needed cash to pay rent, to buy food.” Over a stretch of months she cycled through a churn of short contracts across multiple platforms. Wired's own framing of her experience described roughly twenty contracts across five different platforms, the work soul-crushing and unstable in equal measure. In her account, projects vanished without warning: no message, no explanation, just an abrupt loss of access to the systems she had been working in. She was, as she put it, interviewed by an AI and managed by recent graduates — professionals decades into their careers taking direction from people barely out of school.

Fowler's ambivalence is the point. She knew, even as she did it, that she was helping to build the very systems that could erode her screenwriting career. She did it anyway, because the alternative was not abstract principle versus profit. It was rent versus no rent.

The moral weather around the decision is heavy. Screenwriter Robin Palmer, who has written television movies for the Disney Channel and Hallmark, spoke with CBS News about doing AI training work and acknowledged that some of her peers compared the choice to crossing a picket line. That comparison is not casual in this industry, and it is not deployed lightly. It carries the full weight of one of the most consequential labour actions in recent American history.

Editor Gabe Sena, another entertainment worker who has taken on AI fine-tuning as a side hustle, offered a different framing — one less about morality than about survival inside a shifting field. “I'm mid-career,” he told The Hollywood Reporter, “and I don't want to be a dinosaur in my field.” Steven Woolworth, a former development executive who had worked with HBO and Studio TF1 America, found AI work through the platform Mercor after a year and a half without a job. The pattern is consistent across the reporting: these are not dilettantes or technophiles. They are skilled professionals who exhausted the traditional market and found the door to AI training propped conveniently open.

Even those who have not had to make the choice decline to condemn those who do. Vince Gilligan, the creator of Breaking Bad and the more recent Pluribus, told the publication that he had been extraordinarily lucky throughout his career and would not judge people simply trying to provide for their families. That refusal to judge, from someone with the standing to judge, is itself a kind of data point. It tells you that inside the industry, the people doing this work are not regarded as villains. They are regarded as casualties.

What the work actually is

The technical term for the labour at the centre of this story is Reinforcement Learning from Human Feedback, or RLHF. It is one of the least-understood and most consequential processes in modern AI, and it is worth dwelling on, because the entire paradox of this moment lives inside its mechanics.

A large language model, in its raw pre-trained state, is a statistical engine for predicting the next token in a sequence. It has ingested an enormous corpus of human writing and learned the patterns within it, but it has no inherent sense of what makes an answer good, useful, tasteful, funny, or true. Left to its own devices it will produce text that is fluent and frequently useless. RLHF is the process by which human judgement is injected back into the system. Humans are shown the model's outputs and asked to rank them, score them, correct them, and explain why one response is better than another. Those human preferences are distilled into a reward model, which is then used to nudge the underlying system toward outputs that humans actually prefer.

In other words, RLHF is the stage at which raw machine fluency gets converted into something that feels human. And the people best qualified to perform that conversion — for creative tasks specifically — are precisely the people who have spent careers cultivating taste: the writers who know why a line of dialogue works, the editors who know where a scene should breathe, the executives who know why a pitch is dead on arrival. The system needs their judgement. It cannot manufacture taste on its own. So it rents theirs.

The economics of this are not generous, but they are not nothing. The Hollywood Reporter noted that creative writers could earn up to $44 an hour to work on an AI project through the jobs platform Handshake, while a music professional with a master's degree or above could command up to $100 an hour. Tim Friedlander, a voice actor and a leader at the National Association of Voice Actors, captured the seductive arithmetic without endorsing it, noting that he would not necessarily fault someone offered potentially $1,200 for four hours of work. For a creative worker accustomed to feast-or-famine income, those rates can look like a lifeline. They also sit, in many cases, below the worker's true market value when traditional work was available — which is part of what makes the bargain feel less like opportunity and more like a managed descent.

The company organising human intelligence

If RLHF is the mechanism, the marketplace is the structure that scales it. And no company better embodies the dynamic than Mercor, the firm through which several of the entertainment workers in the reporting found their gigs.

Mercor was founded in 2023 by three young entrepreneurs — Thiel Fellows, recipients of the $100,000 grants that billionaire investor Peter Thiel awards to people who forgo or abandon college to build companies. By their own account the founders set out to organise human intelligence to power the AI economy, and the business they built does precisely that. Originally conceived as a hiring company that assessed candidates by analysing interview transcripts, résumés, and portfolio sites, Mercor pivoted after it found itself sitting on a network of specialised experts. It now recruits highly skilled professionals — including writers, scientists, doctors, and lawyers — to provide the human feedback that trains AI models.

The scale of the operation is the part that should make anyone pause. In October 2025, Mercor raised $350 million in a Series C round led by Felicis, with participation from Benchmark, General Catalyst, and Robinhood Ventures, quintupling its valuation to roughly $10 billion. Its early backers reportedly included Thiel himself, the Twitter co-founder Jack Dorsey, and former US Treasury Secretary Larry Summers. The company has said it pays out more than $1.5 million per day to contractors, with tens of thousands of experts on its roster earning, on average, north of $85 an hour.

Read those figures back to back and the architecture of the moment snaps into focus. A company valued at ten billion dollars, founded by people in their early twenties, has built a profitable business out of channelling the accumulated expertise of mid-career professionals — many of them displaced from their own fields — into the refinement of systems that compete with those same fields. The revenue model, as more than one analyst has observed, is not merely selling software. It is monetising the slice of the roughly five-trillion-dollar US knowledge-work economy that AI hopes to absorb. The displaced expert is not incidental to the business. The displaced expert is the supply chain.

The numbers behind the squeeze

It would be easy to dismiss all of this as a handful of anecdotes dressed up as a trend. The data does not allow that.

The jobs platform Indeed has tracked a striking shift in the arts category specifically. Between May 2025 and April 2026, the share of arts-related postings that referenced AI roughly doubled, from about five per cent to about eleven per cent. That alone might be unremarkable in an economy saturated with AI hype, except for the comparison: across all sectors combined, the share of AI-related postings rose from 2.8 per cent to 5.5 per cent over the same window. The arts, in other words, are absorbing AI-adjacent work faster than the economy as a whole. The field most romantically associated with irreplaceable human originality is the field tilting most steeply toward training its mechanical understudy.

Meanwhile, the traditional execution work is genuinely thinning. A widely cited analysis published in November 2025 by Bloomberry, which examined roughly 180 million global job postings, found that the picture for creative occupations is bifurcating. Strategic and directorial roles — creative directors, creative managers, creative producers — have held up reasonably well. Execution roles have not. Computer graphic artist postings, the analysis found, fell for two consecutive years: down twelve per cent in the first year of the window and a further, steeper drop in the second. The pattern is consistent with what the entertainment workers describe from the inside. The work that can be specified, scoped, and reproduced is being squeezed. The work that requires judgement about what should be made at all is, for now, more insulated.

That bifurcation is worth holding onto, because it complicates the simplest version of the story. AI is not flatly destroying creative labour. It is reorganising it — pulling value upward toward a narrowing band of strategic roles while draining it from the broad middle where most working creatives have always made their living. And it is precisely that broad middle that is now being recruited, at below-market rates, to refine the systems doing the draining.

The implicit bargain, broken and rebuilt

There is a useful way to think about this that predates the current AI boom. Every cultural worker who ever put work into the world participated in an implicit bargain: I make something, it enters the commons, and over time it shapes the people who come after me, including those who will compete with me. That is how culture has always renewed itself. The young writer learns by absorbing the old writer's scripts. The apprentice editor studies the master's cuts. Influence flows forward, and the people who shaped the field are gradually succeeded by the people they shaped.

Generative AI breaks that bargain in two distinct ways, and it is important not to conflate them.

The first break happened during pre-training, and it happened without consent. The models that now compete with creative workers were built, in significant part, on the accumulated output of those same workers — the scripts, the films, the articles, the images scraped from the open internet and ingested as training data. Storyboard artist Phil Langone, who has worked on major productions, told The Hollywood Reporter in unambiguous terms that AI had crawled his own work and stolen from him. The first break, in other words, took the worker's past output without asking.

The second break is the one this story is really about, and it is subtler and arguably more troubling. Having absorbed the worker's past output, the systems now require the worker's present judgement to become genuinely good. And this time they are asking — they are even paying. But they are asking under conditions that strip the work of everything that made creative labour meaningful: no authorship, no credit, no residuals, no ownership of the improvement you create, and a nondisclosure agreement ensuring you cannot even say what you did. You contribute the most refined product of your career — your taste — and it disappears into a model that bears your fingerprints nowhere and competes with you everywhere.

This is the heart of the paradox the reporting circles. The creative workers who helped build the cultural knowledge base that AI trained on are now being paid, below their market value, under NDAs, with no job security, to refine the same systems. The bargain that once renewed culture has been industrialised and inverted. Influence no longer flows from master to apprentice across a shared field. It flows from worker to machine across a one-way valve, metered by a platform, and the worker is paid a contractor's wage to operate the valve that drains their own profession.

Training your replacement is an old story

It helps, here, to resist the temptation to treat this as unprecedented. Workers have been asked to train their replacements before, and the history is instructive precisely because it is so unglamorous.

In the 1990s and 2000s, the practice became notorious in American information-technology departments, where employees were sometimes required, as a condition of their severance, to train the lower-cost offshore or visa-holding workers who would take over their roles. The cruelty of the arrangement was widely reported and widely resented: you not only lost your job, you were made to hand over the institutional knowledge that made the job possible, often under threat of forfeiting your exit package if you refused. The phrase “training your replacement” entered the language as a byword for managed humiliation.

What is happening now is both like and unlike that precedent. It is like it in the basic structure: skilled workers transferring their hard-won expertise to whatever will perform their function more cheaply. It is unlike it in two respects that cut in opposite directions. On one hand, the AI worker is not usually being coerced by their current employer as a condition of severance; many, like Fowler, walk in voluntarily because the traditional market has already collapsed beneath them. On the other hand, the replacement they are training is not a person who will go on to have their own career, raise a family, and eventually train someone else. It is a system that can be copied infinitely, never tires, never unionises, and never ages into obsolescence the way a human does. You are not training a successor. You are training a substitute that does not succeed you so much as it absorbs you.

That distinction matters for the long-term question, because the old offshoring story still left the work in human hands somewhere. The new version aims to remove it from human hands altogether — and asks the very humans being removed to make the removal smoother.

Ghost work, made visible

The entertainment-industry version of this story is genuinely new in its glamour and its visibility, but the underlying labour pattern is not. The anthropologist Mary L. Gray and the computer scientist Siddharth Suri gave it a name in their 2019 book: ghost work. They were describing the vast, deliberately invisible human workforce that makes automated systems appear seamless — the people labelling images, moderating content, sorting data, and correcting errors behind the curtain of apparent machine intelligence. Their findings were stark: this work typically pays below the legal minimums of traditional employment, comes with no benefits, and can be terminated at any moment for any reason or none. The entire design intent of ghost work is for the customer to believe the magic happened automatically, when in fact a human was always in the loop, paid as little as possible and kept out of sight.

What the Hollywood story reveals is ghost work climbing the prestige ladder. The people now doing it are not anonymous crowdworkers in the developing world — the population Gray and Suri largely documented — but credentialed professionals from one of the most visible industries on earth. And yet the structural features are identical: instability, invisibility, no benefits, abrupt termination, and an NDA to ensure the curtain stays drawn. Fowler's description of being hired and fired across projects with no explanation, of vanishing access and silent dismissal, reads almost as a direct transcription of the ghost-work conditions catalogued years earlier. The difference is the calibre of the labour being burned. The system has discovered that for the hardest creative judgements, it needs not just any human in the loop but a highly trained one — and it has built a market to extract that judgement on ghost-work terms.

The legal system is beginning, haltingly, to notice. Scale AI and Surge AI, two of the largest platforms in the RLHF and data-labelling economy, have each faced lawsuits alleging worker misclassification — the practice of treating workers as independent contractors while controlling their hours, tools, and methods closely enough to look a great deal like employment. The complaints describe unpaid training, unrealistic deadlines, and pay structures that erode the headline rate. More disturbingly, Scale AI has faced litigation alleging that contractors were exposed to deeply traumatising content — material involving suicide, violence, and child sexual abuse — while training systems for major technology firms, and that promised mental-health support never materialised. The creative workers training entertainment models are mostly insulated from the worst of that content. But they are exposed to the same underlying legal and structural ambiguity: a workforce that does indispensable work for trillion-dollar ambitions while being classified, paid, and protected as if it were doing something marginal.

The scab question, reconsidered

Return now to Robin Palmer's peers, and their comparison of AI training to crossing a picket line. The metaphor is worth taking seriously rather than waving away, because the 2023 Writers Guild of America strike is the unavoidable context for everything happening now.

That strike — 148 days, around 11,500 screenwriters, one of the longest work stoppages in the guild's history — was, more than any prior Hollywood labour action, a fight about artificial intelligence. The writers won meaningful protections: AI could not be credited as a writer, AI-generated text could not be treated as literary material, and studios could not use AI to undercut writers' pay or compel them to build on machine-generated drafts. It was a landmark, and it established that organised creative labour could extract real concessions from capital on the AI question.

But the strike's protections, by design, govern the relationship between writers and the studios that employ them. They do not govern the relationship between a freelance writer and an AI-training platform operating entirely outside the guild's jurisdiction. A writer who takes an RLHF gig is not violating their WGA contract. They are stepping into a parallel labour market the contract never reached. That is the precise sense in which the picket-line metaphor both fits and fails. It fits because the writer is, in aggregate, strengthening the technology the guild fought to contain. It fails because there is no picket line to cross — no organised action, no struck employer, no solidarity structure to betray. The work happens in a jurisdictional vacuum the union has not yet figured out how to enter.

That vacuum is the real scandal, and it is structural rather than personal. To frame the individual worker as a scab is to demand heroic sacrifice from the most precarious people in the system while the comfortable look on. The studios that scrape the work, the platforms that broker the labour, and the investors who collect the returns face no such moral scrutiny. It is the unemployed writer choosing between an AI gig and an empty refrigerator who gets cast as the traitor. There is something badly miscalibrated about a moral economy that lands its hardest judgement on its weakest participant. The more honest target is the absence of any collective framework — any union, any regulation, any floor — that would let creative workers refuse this work without starving, or do it on terms that did not strip them of dignity.

The cultural-ecology problem

All of which brings us to the question the Hollywood Reporter investigation poses but cannot answer, the one that outlives any individual's rent payment. For the mid-career worker, taking the gig is a financial decision, full stop, and no one is entitled to second-guess it. But at scale, across thousands of such decisions, something larger is being decided, and it deserves to be named clearly.

A creative culture renews itself through a particular and fragile process. People make things. Other people respond, absorb, react against, and build something new. Taste is transmitted not through specification but through immersion — through years of doing the work, getting it wrong, watching what lands and what dies, and slowly developing the judgement that cannot be written down because it lives in the body and the ear. That judgement is the actual asset. It is what the AI systems lack and what they are paying displaced experts to supply.

Here is the cultural-ecology problem in its sharpest form. The judgement being harvested through RLHF was cultivated under conditions that no longer exist. Fowler and her peers learned their craft inside a functioning industry — staffed writers' rooms, real productions, mentorship, the accumulated experience of getting paid to make things and learning from the making. They are now feeding that hard-won judgement into systems whose growth is correlated with the disappearance of exactly those conditions. If the execution-level work continues to thin, where does the next generation acquire the taste that this generation is busy uploading? You cannot develop a screenwriter's instinct for dialogue by scoring an AI's attempts at it under an NDA. You develop it by writing dialogue that real people perform for real audiences, over and over, for years.

This is the deeper paradox, the one beneath the visible one. The visible paradox is that creative workers are refining the systems that threaten them. The deeper paradox is that the resource those systems most need — cultivated human taste — is non-renewable under the conditions the systems create. The models are, in effect, drawing down a reservoir of judgement that was filled by an industry they are helping to drain, and there is no obvious mechanism by which the reservoir refills. A culture can run for some time on its accumulated stock of expertise. It cannot run forever on a stock it has stopped replenishing.

None of this means the machines will produce work indistinguishable from the human article, or that human creativity is about to vanish. The Bloomberry data itself suggests a more textured outcome, with strategic and directorial judgement holding firm even as execution erodes. The likeliest near future is not the abolition of human creativity but its concentration — a smaller number of people at the top of the value chain, directing systems trained on the harvested judgement of a much larger number of people who have been pushed out of the middle. That is not the end of culture. It is the enclosure of it.

What would have to change

If there is a forward-looking answer here, it is not the consoling one, and it is not a prediction that the machines will fail. They may not fail. The honest answer is that the terms of this bargain are not laws of nature; they are choices, and they could be made differently.

The choices that matter are mostly structural. The misclassification suits against Scale AI and Surge AI hint at one front: if the people performing RLHF were treated as the employees their working conditions suggest they are, the economics of the entire data-labour market would shift, and the most exploitative edges would dull. The WGA's 2023 victory hints at another: collective bargaining reached the studio relationship, and there is no reason in principle it could not eventually reach the platform relationship, if creative labour organises across the jurisdictional gap that currently swallows this work. And the unconsented foundation of the whole edifice — the scraped past output that storyboard artists and writers describe as theft — remains live in courtrooms, where the question of whether training on creative work without permission is lawful has not been settled.

The deepest choice, though, is about the reservoir. If cultivated human taste is the scarce input that makes these systems valuable, then a culture that wants to keep producing distinctively human work has an interest — an economic interest, not merely a sentimental one — in keeping the conditions that produce that taste alive. That means defending the broad middle of creative employment, not just the strategic peak. It means recognising that the people best qualified to teach machines how to be human are valuable precisely because they remain something the machines are not, and that paying them ghost-work wages to disappear into a model is a way of spending down a resource no one knows how to remanufacture.

The veteran writer's question — how many times can you tell a machine it's wrong without losing your mind — turns out to be the right one, just aimed slightly off. The real question is not how much the individual can endure. It is how long a culture can keep asking its most skilled people to pour their judgement into systems built to make that judgement redundant, before it discovers that the well it has been drawing from was never going to refill itself. That is not a problem any single worker at any single laptop can solve. It is a problem about what kind of bargain a society is willing to offer the people who make the things it loves. Right now, the bargain on the table is rent in exchange for obsolescence, paid by the hour, under an NDA. The people taking it are not the ones who should have to answer for it.

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. Ruth Fowler, “I Work in Hollywood. Everyone Who Used to Make TV Is Now Secretly Training AI” (republished excerpt), Longreads, 12 May 2026. https://longreads.com/2026/05/12/training-ai-tv-industry/
4. “In Hollywood, AI gig work could be the 'new waiting tables,'” Marketplace, 20 May 2026. https://www.marketplace.org/story/2026/05/20/in-hollywood-ai-gig-work-could-be-the-new-waiting-tables
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. 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-350m-series-c/
7. “Mercor,” Wikipedia. https://en.wikipedia.org/wiki/Mercor
8. “I analyzed 180M jobs to see what jobs AI is actually replacing today,” Bloomberry, 3 November 2025. https://bloomberry.com/blog/i-analyzed-180m-jobs-to-see-what-jobs-ai-is-actually-replacing-today/
9. “Scale AI Sued for Subjecting Contract Workers to 'Depraved' Content,” Inc., 2025. https://www.inc.com/sam-blum/scale-ai-sued-for-subjecting-contract-workers-to-depraved-content/91111116
10. “Training Artificial Intelligence and Employer Liability: Lessons from Schuster v. Scale AI,” Epstein Becker Green / Workforce Bulletin, 2025. https://www.workforcebulletin.com/training-artificial-intelligence-and-employer-liability-lessons-from-schuster-v-scale-ai
11. “Surge AI and Scale AI face lawsuits over alleged worker misclassification and unpaid wages,” Complete AI Training. https://completeaitraining.com/news/surge-ai-and-scale-ai-face-lawsuits-over-alleged-worker/
12. “What Companies Can Learn from the Surge AI and Scale AI Lawsuits,” Worksuite. https://worksuite.com/resources/insights/ai-misclassification-lawsuits
13. Mary L. Gray and Siddharth Suri, Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass, Houghton Mifflin Harcourt, 2019. https://ghostwork.info/
14. “2023 Writers Guild of America strike,” Wikipedia. https://en.wikipedia.org/wiki/2023_Writers_Guild_of_America_strike
15. “What the WGA Contract Tells Us About Workers Navigating AI,” Tech Policy Press. https://www.techpolicy.press/what-the-wga-contract-tells-us-about-workers-navigating-ai/

---

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...