Last year, a San Francisco-based AI company called Mercor exploded onto the tech scene, just as US workers were buckling under the grip of unemployment.

The online job marketplace connects contractors — often struggling unemployed workers — with companies like OpenAI, the creator of ChatGPT. The gigs have a ghoulish tint, since the basic idea is for former employees to teach AI how to do their old jobs. Adding insult to injury, Mercor is known for treating its workers poorly.

Now, a mini documentary published by More Perfect Union pulls the curtain back even further on companies like Mercor. In the video, reporter Karen Hao interviews a series of data workers — many of whom wanted to remain anonymous for fear of retaliation — about what it’s like to be employed by what the nonprofit calls “America’s AI sweatshops.”

Perhaps the project’s most eye-catchingly tragic stat comes from 2025 research by the Communication Workers of America, which found that among these tenuous workers training AI systems, a staggering 22 percent said they’d experienced homelessness due to their meager wages.

I Tracked Down the Hidden Workers Secretly Powering ChatGPT

In the documentary, Hao also cites startling data from a study led by labor researcher Tim Newman. According to his piece, about 86 percent of data workers — those who may be training the AI models you use every day — struggled to pay their bills last year. Nearly one-quarter relied on public assistance programs, such as food stamps and Medicaid.

One interviewee who went by the pseudonym Jen told Hao that she faced an uphill battle in the job market after graduating from an Ivy League school with a PhD more than a year ago. Without any promising career leads, she was forced to move in with her sister and rely on food stamps. Out of desperation, she applied for a gig opportunity from Mercor that offered $55 an hour, far surpassing what she was making as a cashier and substitute teacher.

“I think the role I saw was philosophy intelligence analyst,” Jen told Hao. “I’m looking and I’m like, ‘Well, why wouldn’t I be able to do that?'”

But she soon realized there was reason for skepticism. A mere two weeks after her first project, Mercor entirely pulled the rug out from Jen entirely.

“We all get a message in our group comms where it’s like, ‘Actually, like, this contract is ending,'” Jen explained about her experience.

It turned out there was reason for her to be skeptical — and Mercor’s reported workforce of 30,000 may want to take note.

More on AI labor: AI Companies Are Treating Their Workers Like Human Garbage, Which Be a Sign of Things to Come for the Rest of Us