Hollywood's Steadiest Paycheck: $200 an Hour to Train Its Replacement

TL;DR

On 22 August 2026, The Guardian reported that award-winning Hollywood writers, directors and producers are taking gig work at AI training marketplaces, paid $12 to $200 an hour to teach models how to schedule a two-day shoot, assemble a pitch deck, and pull clean speech out of a noisy Little League game. The middlemen are Mercor, micro1 and Handshake; the buyers include Anthropic and OpenAI. What they are producing is not scraped text. It is expert-authored evaluation tasks and reinforcement-learning environments. If you have ever wondered where an agent's domain competence actually comes from, this is the receipt, itemized by the hour.

The job is teaching the model your job

Ruth Fowler, who created the BBC One drama Rules of the Game and co-wrote Paramount Plus series Little Disasters, told the Guardian she signed up "because I was thinking: wow, I'm always broke." Her tasks were not abstract. She trained a model to produce a detailed schedule for a hypothetical two-day shoot: which filming permits were required, what hazards the location carried, what the daylight would do, who needed to be on set for each shot, and what child-protection provisions applied. The inputs were the same messy artifacts a real production runs on, including emails, a shooting script and production paperwork.

A Los Angeles documentary director, who asked not to be named, was hired to teach a model to transcribe video accurately. He worked through footage of a Little League baseball game where crowd noise and music smeared the dialogue, wrote a visual description of each speaker so the system could tell them apart, and characterized their accents. His summary of the arrangement: "I was essentially handed a shovel and asked to dig the grave of my profession."

The unusual part is not the fatalism. It is the paper trail. This pipeline is normally invisible, buried under a phrase like "trained on a large and diverse corpus." This time it filed a timesheet.

The rate card is public, and it is not a secret at all

Mercor's front page lists open engagements with hourly rates attached. As of today it is advertising an FP&A expert at $80 to $90 an hour, product management at $80 to $90, accounting at $70 to $80, design at $70 to $80, architecture at $75, sales and marketing at $65 to $90, and a healthcare insight study at $120.

The Guardian reports micro1 was advertising for people with "demonstrated mastery in film, television, digital, or live event production" and five or more years of credited experience, to "design and author evaluation tasks that simulate realistic production-management scenarios, such as budget reconciliation, scheduling adjustments, and vendor or crew coordination," at up to $85 an hour.

Read that listing again. The deliverable is not an answer. It is a task, plus the standard by which an answer gets judged.

What your production notes turn into

micro1 names its own product lines on its site, and they say the quiet part out loud: Realm is "frontier evaluations and RL environments for AI Labs," Cortex is "the visibility and improvement layer for agentic AI," and its robotics line sells "high quality expert-demonstrated data for frontier robotics labs."

A reinforcement-learning environment is less like a textbook and more like a driving test. You do not hand the model the answer. You hand it a scenario, let it try, and score the attempt against a rubric. The scenario and the rubric are the whole asset, and neither can be scraped off the open web, because nobody publishes their shooting schedule with a grading key attached. So the examiner is now a line producer with twenty years of credits and an invoice.

This market is not small. Mercor crossed $2 billion in gross annualized revenue in June 2026, roughly double its pace at the start of the year, per Dealroom. Forbes reported in July that the company was in talks to raise about $500 million at a $20 billion valuation, after raising $350 million at $10 billion in October 2025. Note the word "gross": that figure is closer to payment volume than to net revenue, since the contractors keep most of it. The margin is thinner than the headline. The volume is the point.

Why the supply of willing experts exists

Because the alternative work is gone. FilmLA's year-end research report put Greater Los Angeles at 19,694 on-location shoot days in 2025, down 16.1 percent from 23,480 in 2024, with every category sitting well below its five-year average.

The national employment picture matches. Bureau of Labor Statistics figures put motion picture and sound recording employment at 321,700 in June 2026, down another 3,600 jobs that month, against a post-pandemic peak the Guardian puts near 450,000 in July 2022. Meanwhile Netflix told investors that generative AI touched roughly 300 titles in the first half of 2026, and Ted Sarandos said the docuseries The American Experiment carries 17 minutes of AI-enhanced footage made "twice as fast and at half the cost" of the alternative.

So the labor pool is idle, the buyer is flush, and the price clears. That is not a conspiracy. It is a market doing exactly what a market does, which is somehow more unsettling.

What a builder should actually take from this

Three things, and none of them are about Hollywood.

1. Domain competence is manufactured, not emergent

When a frontier model suddenly gets good at a specialized workflow, the likeliest explanation is that somebody paid practitioners of that workflow to write task and rubric pairs. If you want an agent that is genuinely good at your niche, the frontier labs' own method is available to you at small scale: stop collecting transcripts and start writing graded scenarios with the person who does the job.

2. The rubric is the moat, not the corpus

Fowler's shoot schedule is worth money because it encodes what "correct" means when the answer is a plan rather than a fact. Permits, hazards, daylight, child-protection rules. That judgment is what nobody publishes, and it is what your eval suite is missing.

3. Capability follows labor-market slack

Watch which professions get modeled first. It will not be the most important ones. It will be the ones where skilled practitioners are simultaneously expert, idle, and reachable. Hollywood is the current example because a 16 percent production decline made a lot of very good line producers available on a Tuesday afternoon.

The caveats, stated straight

• The $12 to $200 band spans different agencies and task types. It is a range, not a going rate, and the low end is real.
• Mercor's $2 billion is a gross annualized run rate, closer to payment volume than net revenue. The $20 billion valuation was reported as in talks, not closed.
• The FilmLA and BLS declines began before generative AI was widely used in production. Streaming pullback, the 2023 strikes and runaway production drove most of it. AI is arriving into a slump it did not start.
• The practitioners themselves are unconvinced about quality. One writer with credits at Amazon, HBO, NBC and Lionsgate told the Guardian the models are "no closer to approximating human emotion" and could write an episode of CSI but not Severance.
• Nothing here says these specific tasks trained any specific shipped model. The agencies contract with multiple labs and disclose nothing about which capability landed where.

Key Takeaways

• The training-data pipeline has a public price list. Mercor is advertising $65 to $120 an hour for professional expertise, and Hollywood creatives report $12 to $200 for production and screenwriting work.
• The product is evals and RL environments, not text. micro1 sells "frontier evaluations and RL environments" by name. The scarce asset is the grading standard, not the transcript.
• The market is large and growing fast. Mercor hit a $2 billion gross annualized run rate in June 2026, roughly double its January pace, though contractors keep most of that.
• Supply comes from a genuine industry slump. Greater LA logged 19,694 shoot days in 2025, down 16.1 percent, and US motion picture and sound recording employment sits at 321,700.
• Copy the method, not the mood. If you want an agent good at your domain, hire the practitioner to author graded scenarios. That is what the labs are doing, in public, by the hour.

Sources: The Guardian, "Digging the grave of my profession": the Hollywood creatives training AI to do their jobs, 22 August 2026, Mercor public listings, micro1, Handshake AI, Dealroom, Mercor doubles to $2B gross revenue run rate, Forbes, Mercor in talks at $20 billion, 9 July 2026, FilmLA Q4 and year-end 2025 research report, 15 January 2026, BLS, Motion Picture and Sound Recording Industries (NAICS 512), TheWrap, June 2026 industry jobs update, Cartoon Brew, Netflix says genAI was used on around 300 titles in 2026 so far