AI Training Boom Masks Deeper Job Displacement Fears

In a nondescript office in downtown San Francisco, a company called Mercor has assembled an army of more than 100,000 freelancers—winemakers, music producers, physicians, lawyers—to teach artificial intelligence systems how to do what humans have spent years perfecting. It is, on its surface, a boom. Mercor's CEO, Brendan Foody, was 23 when his company became a multibillion-dollar enterprise. Four years ago, when ChatGPT arrived, Foody and two high school friends recognized that to improve, AI would need far more training data than the public internet could provide. So they built a marketplace. When OpenAI or Anthropic wants to make their models smarter in a particular field—finance, architecture, poetry, medicine—they hire Mercor's freelancers to feed those systems their expertise. The work pays between $20 and $200 an hour. It is, Foody insists, a career of the future.

But the workers themselves tell a different story. Shawna Miller, a winemaker, Robbie Hiser, a music producer, and Dr. Melania Poonacha, a physician, all do this work on the side. They have full-time jobs. When asked if they worry they are training themselves out of employment, they say no—that human creativity, lived experience, and judgment cannot be replicated by machines. Hiser points out that an AI has never had a girlfriend break up with it, has never lived through heartbreak, and therefore cannot write a song that captures that truth. Yet even as these workers express confidence, the hours are unpredictable. Many of the more than a dozen trainers interviewed said the gig work comes sporadically, if at all. There is no health insurance, no retirement, no stability—the things that workers in previous eras, even in the same fields, took for granted.

The optimistic case is being made everywhere. Kim Herlihy, a trial lawyer at the Columbus, Ohio firm Vorys, Sater, Seymour and Pease, works with an AI trained on the expertise of her firm's senior partners. The system helps her draft opening statements, offering feedback on legal phrasing. She does not believe AI will walk into a courtroom and replace her. A lawyer's work, she argues, is fundamentally human—clients need to look their attorney in the eye and ask for advice. Brendan Foody echoes this. He foresees a utopia in which AI handles the dull, repetitive work, freeing humans to be more productive and creative. When asked directly if AI training is simply a way to replace workers, Foody says the work is evolving, not disappearing. There will always be new things for humans to do.

But Clara Shih, who until January oversaw major AI divisions at Salesforce and Meta, quit her position to found a nonprofit helping young people navigate this new economy. She has become the industry's most visible skeptic. She has watched AI reduce teams of dozens to teams of one or two. She has seen what happens when a technology arrives faster than society can absorb it. Her family came to the United States from Hong Kong in the 1980s and settled in Ohio, where globalization was hollowing out factories and the communities that depended on them. She sees the same pattern unfolding now. The tasks that companies traditionally assign to young people—market research, drafting memos, initial analysis—are precisely what AI does best. Entry-level workers are already feeling the squeeze: hiring in AI-exposed fields like software development has slowed, and wages have declined. Nearly three-quarters of Americans now fear AI is coming for their jobs.

Daron Acemoglu, a Nobel Prize-winning economist at MIT who studies technology's impact on labor, does not mince words. When Foody and others claim that new jobs will emerge to replace those lost, Acemoglu calls it nonsense. The number of people employed in AI training is small relative to the number who will be displaced by automation. That is the entire point of automation. On the current trajectory, he says, unemployment could triple within a decade. The Industrial Revolution's first phase lasted 80 years. This one is happening in one or two years, across many sectors simultaneously. When asked if there are things humans do that machines cannot, Acemoglu agrees—today. But in 15 years, he does not know what AI will be capable of.

The tech industry has begun to sense the public's alarm. Nearly three-quarters of Americans fear for their jobs. In response, companies like OpenAI have shifted their messaging. Sam Altman, OpenAI's CEO, recently said the industry underestimated how important it would be to keep people at the center of the economy. But Shih is skeptical of this pivot. The AI labs, she argues, have every incentive to sell their products to the public now. That does not mean they believe their own reassurances about job safety. Acemoglu offers a path forward: design AI systems that make humans better rather than expendable, restructure the tax code to incentivize hiring real people, and begin the difficult conversation about what society wants from this technology before the window closes. But he is not optimistic. The future, he says, is subject to our choice and agency. The question is whether we will choose to act.