When the people training AI are training themselves out of a job

The work itself has shifted toward higher-credentialed, more specialized expertise. Earlier data-labeling operations relied on low-paid overseas workers tagging images. Today’s AI labs need mathematicians to annotate formal proofs, lawyers to mark up legal briefs, professors to grade essays. The open internet has largely been scraped already, so further gains now depend on capturing the tacit expertise that never made it online in the first place.

Some firms are going further still, trying to recreate entire workplaces in miniature. Mercor’s founder, Brendan Foody, described the acquisition of Deeptune, a startup that builds simulated versions of workplace software like Slack and Salesforce, as solving “the bottleneck to a frontier lab automating everything that people do.” In practice, that means pairing role-players as clients and employees to model exactly what happens inside, say, an investment bank.

What the work actually looks like

The daily reality of these jobs is less glamorous than the funding rounds suggest. Reporting by The Guardian in 2025 described AI “raters” contracted through Hitachi’s GlobalLogic to evaluate and moderate output from Google’s Gemini and AI Overviews. One rater described a task timer that shrank from 30 minutes to 15, in which she was expected to read, fact-check, and rate roughly 500 words of AI response each time. Rachael Sawyer, a Texas-based technical writer who has rated Google’s AI products since March 2024, said she was expected to complete dozens of such tasks a day, each within a 10-minute window, and that the job pulled her into content moderation, including flagging violent and sexually explicit material, that was never mentioned in her onboarding or job description. Researcher Adio Dinika, who studies the human labor behind AI at the Distributed AI Research Institute, put it starkly: “AI isn’t magic; it’s a pyramid scheme of human labor.”

The professionals doing more specialized work described a similar arc. Amanda Brown, an assistant professor of biology at Tarleton State University in Texas, took on gigs through Mercor and Handshake and told The New York Times the pay looked appealing at first, but the hours ballooned, deadlines tightened, and feedback from supervisors felt demoralizing. Carolina Perez Sands, a speech and language pathologist who trained language models in Portuguese and creative writing for Mercor, described a similar trajectory to The Wall Street Journal’s “The Journal” podcast in June 2026. Her corrections became unnecessary within about a week as the model absorbed them, and she eventually concluded that “my job is actually making this more of a monster.” She has since left the industry for a sales role.

Workers training themselves out of a role within months came up across multiple accounts. It’s also, by design, the business model. These firms need AI models to keep improving to prove their value to clients, but they need those same models to stay imperfect enough that clients keep coming back for more data.