AINEWS
Specialist agencies pay Hollywood writers and producers $12 to $200 an hour to train Anthropic and OpenAI models on the craft they are being priced out of.
Ruth Fowler spent an afternoon showing an AI how to plan a two-day shoot in the Hollywood Hills. She walked it through the child protection rules, the daylight windows, the permit lead times, and the small contingencies that decide whether a day of filming exists at all. The buyer was a specialist training agency whose hourly rates for senior creative work run as high as $200, and whose other clients, by Fowler's account, include the biggest names in frontier AI.
The contradiction sits inside the same hour of work. The screenwriter and producer of Rules of the Game (BBC One) and Little Disasters (Paramount+) is being paid to teach a model the craft she is being priced out of. "It's teaching it how to take our jobs," she said. "Wow, I'm always broke."
The training agencies named in the work do not advertise their client lists, but the people they hire know who is reading the output. Several hold contracts with Anthropic and OpenAI, two of the largest US AI labs. The work they broker ranges from screenplay formatting and pitch deck assembly to noisy-footage transcription and the kind of schedule triage a first assistant director handles on day one of pre-production.
The pay spread tells its own story. Hourly rates run from $12 for entry-level annotation work to roughly $200 for senior creatives whose judgment is the actual product. The lower end of that range is the work that scales. The upper end is the work that trains the model on what to ask the lower end to do.
Fowler estimates that production on the projects she would normally staff has fallen "35% or something insane." The figure is hers, not an industry study, and it sits inside a wider slump in motion picture and sound recording work in Los Angeles, where AI is starting to take roles from actors, editors and special effects artists.
The pattern is not confined to Hollywood. The same agencies market the same arrangement to finance, health, law and social-work professionals. Each pool of expertise teaches the model that will eventually compete with that pool of expertise, at a rate the pool cannot refuse while the work is being booked.
Netflix disclosed this year that it used generative AI in 300 of its roughly 1,000 titles. Director Ron Howard is among the filmmakers who have started to embrace the technology. The disclosure is a scale marker, not the story; the story is the labor that turns the disclosure from a line item into a working product. That labor is the part of the supply chain that the AI labs do not run themselves.
A second creative in the trade, an LA documentary director who asked not to be named because the agencies enforce NDAs, spent several weeks teaching a model to transcribe muddled Little League footage, the kind of clip where a coach's whistle and a parent's shouting blur into the dialogue. The task is not glamorous and the rate is at the lower end of the spread. The point the work makes is the cross-domain one: every industry has its own version of muddled Little League footage, and every version is now being fed to a model being trained to clean that footage up on its own.
Fowler has run the math past the fatalism. The "shovel and grave" image, of being handed a tool to dig the hole where her profession will be buried, is the comparison one creative used on the record. The harder line, and the one Fowler returns to in conversation, is what happens when the training class ends.
Right now, the agencies need the workers. When the models stop needing the workers, the same agencies will still hold the contracts. The next round of rate negotiation will happen in a market where the buyer no longer requires a steady supply of the people currently doing the work. That is the open question the current pipeline does not answer: who is left holding the shovel when the digging is done.