Hollywood writers are training the AI that could put them out of work

Some of the creative workers who once fought to limit AI’s role in entertainment are now being paid to help improve it.

Writer and showrunner Ruth Fowler spent eight months doing roughly 20 AI training contracts across platforms including Mercor, Outlier, Turing, Handshake, Micro1 and Task-ify.

In an essay posted on Wired, she talks about work that included rating chatbot answers, judging tone, annotating videos, reviewing images and testing whether AI systems could produce unsafe or disturbing material.

The jobs were presented as flexible remote work. Some advertised high hourly rates. But Fowler’s account describes unpaid tests, confusing instructions, sudden removals and projects that could disappear before contractors earned much at all.

The timing matters. Hollywood is still dealing with the aftershocks of the 2023 strikes, streaming cutbacks and a production slowdown that has hit Los Angeles especially hard.

In that environment, AI training has become a fallback for some writers. It is not a glamorous reinvention. It is a way to keep money coming in when fewer shows are being made.

Los Angeles has fewer shoots

The loss of traditional work is visible in the production numbers.

Back in January, The Hollywood Reporter reported that Luminate found a 24 percent year-over-year drop in major scripted film and television projects shooting in Los Angeles.

FilmLA, the region’s main shoot-day tracker, reported that feature film production in Los Angeles fell 16.8 percent, television shoots dropped 14.7 percent and commercial work declined 14.5 percent. Total shoot days fell to 19,694, down from 36,792 three years earlier.

The nonprofit also reported that on-location production in Greater Los Angeles fell 22.4 percent in the first quarter of 2025, reaching 5,295 shoot days. TV drama production dropped 38.9 percent, TV comedy fell 29.9 percent and reality television declined 26.4 percent.

For workers, those figures are not abstract. Fewer shooting days mean fewer rooms, fewer crews, fewer calls and fewer chances to build a year’s income from entertainment work.

California is fighting back

According to The Hollywood Reporter, Gov. Gavin Newsom signed a bill in July of 2025 raising California’s annual film and television incentive program from $330 million to $750 million. Los Angeles Mayor Karen Bass also announced steps to make filming easier in the city, including faster permit reviews, a pilot program for low-impact productions and proposed lower fees at Griffith Observatory.

The state has used tax credits to draw productions with California ties. The Hollywood Reporter cited the Baywatch reboot, a Snoop Dogg biopic and Ang Lee’s Gold Mountain as examples of projects receiving incentives to film in California.

But the competition is broad. Georgia, Louisiana, New Mexico, New York, New Jersey, Canada and the United Kingdom are all fighting for productions. Streaming has also made the business more global. A show can be financed in one place, shot in another and aimed at audiences far beyond the United States.

That leaves Los Angeles workers stuck between long-term recovery plans and short-term bills.

High rates can mislead

AI training can look attractive in that gap.

According to Ruth Fowler, some expert roles were advertised at up to $150 an hour. Other assignments paid less, but still seemed better than many emergency side jobs.

The problem was how little the headline rate guaranteed.

She describes unpaid application tests, long onboarding processes, delayed starts and projects that produced far less work than expected. In one case, a contract expected to provide 20 hours a week for two months ended after only limited work. In another, tasks appeared after weeks of waiting, but technical problems meant some contractors could not access them before the work ran out.

The arrangement was flexible in name. In practice, workers often had to monitor messages, react quickly and compete for tasks before they vanished.

For people already used to Hollywood’s uncertainty, the rhythm may have felt familiar. The difference was that this version happened through Slack channels, onboarding quizzes and disappearing dashboards.

The work can vanish

The AI jobs described by Fowler were not stable replacements for television employment.

They were short contracts organized through online platforms. Workers had to join communication channels, read detailed rules, sign up for payment systems and wait for access. A project could be staffed quickly, then shut down with little warning.

Fowler writes that Slack channels could disappear, documents could become inaccessible and work interfaces could close without a clear explanation.

The tasks themselves varied. Some required creative judgment, such as evaluating chatbot responses or scripts. Others were repetitive, including video annotation or image review.

The common demand was speed. Workers had to follow changing instructions, meet scoring requirements and complete tasks while they were still available.

A useful way to describe the system is digital day labor: Not because the work is unskilled, but because the opportunity can be brief, competitive and controlled by someone else’s schedule.

Labor questions are growing

The legal questions around AI training work are becoming harder to ignore.

Several lawsuits have accused Mercor of misclassifying workers as independent contractors. The complaints pointed to repeated onboarding, retraining, weekly expectations, short-notice assignments and the need to monitor communication channels.

Independent Contractor Compliance reported in June 2026 that a proposed class action in Texas accused Mercor of misclassification on behalf of about 30,000 professionals who train AI models and chatbots.

The Economic Policy Institute has warned more broadly that misclassified workers can lose protections tied to minimum wage, overtime, unemployment insurance, workers’ compensation and the right to organize.

That context matters. The issue is not only whether a platform communicates badly or whether a project ends abruptly. It is also about who carries the risk when work is irregular.

In a traditional job, supervision, scheduling and downtime can create obligations for an employer. In contractor work, much of that burden can be pushed onto the person waiting for the next task.

Human skill becomes data

The sharpest irony is that AI companies need human expertise, but often break it into tiny measurable units.

Writers are valuable because they understand tone, structure, pacing and dialogue. Editors can detect weak reasoning. Specialists can judge whether an answer sounds credible. Visual workers can spot details that automated systems miss.

But Fowler’s account shows that contractors are often expected to follow strict rubrics rather than use broad creative judgment. Small wording choices could be marked wrong. Scores could fall without clear explanation. Access to future work could depend on speed, compliance and availability.

That is a bitter shift for Hollywood writers. Their profession depends on voice and interpretation. In AI training, those skills may be useful only after they are turned into ratings, labels, rankings and corrections.

The work helps machines sound more human, thus making the humans doing the work easier to replace.

No stable replacement

Hollywood has always been an insecure place to work. Writers wait between rooms. Crew members follow productions. Actors patch together income from different jobs.

But the current squeeze is wider than a normal slow year. Los Angeles is losing shoots. Studios are reducing costs. Other states and countries are offering better financial terms. AI companies are hiring creative people, but often through unstable platforms rather than durable jobs.

For some workers, AI training can still be useful. It can pay a bill. It can fill a gap. It can bring in money when another industry has stopped calling.

But it is not a real answer to Hollywood’s employment crisis.

Sources: Wired, The Hollywood Reporter, FilmLA, Independent Contractor Compliance, Economic Policy Institute

This article is made and published by Kenneth Glad, who may have used AI in the preparation