Last year, a freelance journalist named Katya pivoted to content marketing hoping for more stability. Instead, she found AI had already automated much of the work. Desperate, she clicked a LinkedIn job posting that led her to Mercor, a San Francisco startup that would pay her to train the very technology that displaced her.
Katya's story, reported by The Verge, is not unique. It is the defining labor arrangement of the current AI boom: the people who lost their jobs to AI are now the ones teaching it to get better.
A $10 Billion Middleman
Mercor connects domain experts with AI labs that need human feedback to improve their models. Its client list reads like an AI power ranking: OpenAI, Anthropic, and Meta all use Mercor contractors. The company raised $350 million in a Series C last October, quintupling its valuation from $2 billion to $10 billion in just eight months. It pays out more than $1.5 million per day to over 30,000 active contractors.
The pay varies wildly by expertise. A dermatologist helping build healthcare decision-support tools earns up to $250 an hour. A poet refining an AI's grasp of emotional expression gets $150. A social media marketer writing video captions? $45. And the workers on Mercor's data labeling projects for Meta earned $21 an hour before the company cut their pay to $16 in late 2025, a 24% reduction, after abruptly terminating a project called "Musen" that employed roughly 5,000 people.
Mercor's justification for the pay cut was corporate double-speak at its finest: the new project offered "steadier task volumes" and "higher hour caps, enabling greater weekly engagement." Translation: work more hours for less money per hour.
The Workers Know Exactly What They're Doing
The most striking thing about the people doing this work is how clear-eyed they are. Katie Williams, a 30-year-old video editor who spent six months captioning and rating video clips for Mercor, put it bluntly: "I joked with my friends I'm training AI to take my job someday."
Peter Valdes-Dapena, a 61-year-old automotive journalist laid off in 2024, now critiques AI-generated news articles for the company. His rationalization is practical rather than optimistic: "I didn't invent AI and I'm not going to uninvent it. If I were to stop doing this, would that stop it?"
He has a point. The training will happen regardless. But the dynamic is still bleak. Workers on the terminated Meta project described feeling trapped. One said: "We wanted to boycott this but are not in a financial place to do so. We needed to have the guaranteed income, even if it's demoralizing."
The Numbers Don't Add Up for Workers
Mercor's founders are building toward a potential IPO. Scale AI cofounder Alexandr Wang is worth $3.2 billion. Surge AI's CEO Edwin Chen is worth $18 billion. The wealth being generated by AI training is enormous, but it flows almost entirely upward.
Meanwhile, a Carnegie Mellon study found that AI fails at real-world office tasks 70% of the time, and an MIT study estimated that over 20 million Americans hold jobs that current AI could replace, representing $1.2 trillion in wages. Those two facts sit in uncomfortable tension: the technology is not good enough to do the work reliably, but it is good enough to depress the wages of the people who can.
That is the real story here. Mercor is not a villain. It is a symptom of an AI industry that needs human expertise to function but treats that expertise as a commodity. The contractors installing time-tracking software, the poets earning $150 an hour, the data labelers who got a 24% pay cut before the holidays, they are all feeding a machine that values their knowledge but not their livelihoods.
White-collar unemployment hit a four-year high late last year. The people training AI are not doing it because they believe in the mission. They are doing it because the alternatives are worse.