Ruth Fowler wrote a BBC drama. Now she's teaching AI how to write one.

Ruth Fowler created Rules of the Game, a BBC One series starring Maxine Peake, and co-wrote Little Disasters for Paramount Plus. Last month she spent a chunk of her week teaching an AI model how to build a two-day film shoot schedule from raw production documents: permits, location hazards, daylight windows for cinematography, child-protection requirements for the cast. Not a metaphor. Actual task, actual pay, actual model getting actually better at doing her job.

That's the Guardian's reporting on Hollywood creatives taking AI training gigs, and the framing everyone's landing on is the obvious one: irony, grief, a profession digging its own grave with borrowed shovels. It's a good line. One of the anonymous documentary directors in the piece used almost exactly that phrase. But the irony isn't actually the interesting part. The interesting part is what the pay spread tells you about what these AI companies think they're buying.

Fowler and others are earning $12 to $200 an hour through training agencies under contract with Anthropic and OpenAI, working across finance, health, law, and social work as well as entertainment. That's not one job. A $12/hour rate is rote labeling: flag this, tag that, confirm the transcript is accurate. A $200/hour rate is buying someone's twenty years of professional judgment and hoping it transfers into training data before the person runs out of savings and takes a different job. The range is the tell. AI labs have figured out that expert-level correction data is worth paying up for, and they've found a labor pool willing to sell it at a discount, because the alternative for that labor pool is zero.

Why is that pool suddenly available at Hollywood-writer prices? The numbers in the piece explain it without needing much interpretation. LA shoot days fell 48% between 2021 and 2025, per FilmLA Research. US motion picture and sound recording jobs dropped 28%, from 450,000 in July 2022 to 326,000 this past May. Netflix says it used AI in 300 of its 1,000 titles released in 2026. Ron Howard is releasing an AI-enabled documentary feature. None of that is subtle. The industry cratered for reasons that predate AI (the pandemic, the writers' strike, streamers pulling back investment), and AI arrived exactly when the labor market had already produced a large supply of unemployed, skilled, desperate people. That timing is not a coincidence anyone had to engineer. It's just what happens when a contraction and a capability curve overlap.

Here's the counterargument, and I want to take it seriously before I dismiss it: shouldn't these creatives just refuse? One of the Guardian's anonymous sources basically pre-empted this — he said he took the work because "no amount of abstaining will prevent it." That sounds like rationalization. It might partly be. But it's also just correct as a matter of game theory. If Fowler declines the gig, the training agency finds another out-of-work production coordinator to do it, probably for the same money, possibly with less domain expertise, which if anything makes the resulting model worse rather than stopping it from existing. Individual refusal doesn't remove the demand. It just changes whose expertise ends up encoded and who gets paid for it. I don't love that logic, but I can't find the hole in it.

If you're a builder shipping anything that touches RLHF, expert feedback loops, or fine-tuning on domain data, the actionable version of this is worth sitting with: the market for buying expert human judgment as training data is currently priced low relative to its actual value, because the sellers are in a desperate labor market, not a rational one. That won't last. Whatever expertise you're sourcing cheaply this year through a training-data agency is likely to cost more once the current glut of displaced specialists clears — either because they find other work, or because there are fewer of them left to hire.

Further reading

• 'Digging the grave of my profession': the Hollywood creatives training AI to do their jobs — The Guardian — primary source for all figures and quotes in this piece