The teacher is always eventually replaced by the student.
In the shadow of a technological transformation still taking shape, tens of thousands of credentialed professionals are being paid handsomely to transfer their hard-won expertise into artificial minds — a paradox in which the act of working becomes, perhaps, the act of making oneself obsolete. Mercor, a company barely three years old, sits at the center of this arrangement, channeling two million dollars a day to specialists in fields from radiology to philosophy who are teaching AI systems the very knowledge that defines their careers. The endeavor raises one of the oldest questions in the history of labor — what becomes of the craftsman when the craft is mechanized — now posed at the level of the professions themselves.
- A 22-year-old billionaire's startup is paying 30,000 specialists up to $375 an hour to train AI models in their own fields, creating a lucrative but temporary new class of knowledge work.
- The arrangement carries a quiet contradiction at its core: every hour of expertise transferred to a machine may be one hour closer to the moment that machine no longer needs its teacher.
- London's mayor has warned of mass unemployment in finance and creative industries, while a University of Virginia economist cautions that AI could realistically substitute for most functions defining knowledge work.
- Workers on the platform report high pay and flexibility, but labor economists flag a structural trap — trainers hold no protections and are actively eroding their own bargaining power with each session.
- The U.S. data annotation industry, worth $5.7 billion in 2024, is projected to nearly triple by 2030, yet the deeper question — what society does with displaced knowledge workers at scale — remains unanswered.
Brendan Foody was twenty-two when Forbes named him one of San Francisco's youngest billionaires. His company, Mercor, is now at the center of one of the more unsettling arrangements in modern labor: paying highly credentialed professionals to teach AI systems how to do the very jobs those professionals spent years mastering. Founded less than three years ago, Mercor manages teams that train models built by OpenAI and Anthropic across disciplines ranging from consulting to philosophy.
The economics are striking. Mercor distributes roughly two million dollars daily among approximately thirty thousand specialists. Average hourly rates exceed ninety-five dollars, and in high-demand fields like radiology, trainers can earn up to three hundred seventy-five dollars an hour. The work is project-based, typically lasting several weeks with no guarantee of continuation — yet it draws professionals who value both the flexibility and the income, even as many sense the deeper implication of what they are doing.
Foody frames the moment as the birth of a new category of work, one organized around managing and refining AI agents. He acknowledges that jobs will disappear but argues new categories will emerge to replace them. Critics are less sanguine. London's mayor has warned of mass unemployment across finance, professional services, and creative industries. Anton Korinek of the University of Virginia told the Financial Times that these technologies could substitute for many of the functions defining knowledge work, offering a pointed summary: 'The teacher is always eventually replaced by the student.'
Among Mercor's contractors, reactions range from enthusiasm to measured pragmatism. One consultant described stress-testing AI systems with disaster scenarios — evaluating how models handle damage control and stakeholder management — and refining them based on their responses. Research from Oxford Economics found that forty-one percent of AI trainers hold advanced degrees, and ninety-four percent combine the work with other employment or study.
Labor economist Zoe Cullen of Harvard Business School identifies the structural problem plainly: trainers have no protections against the models they are building, and by sharing their expertise, they are diminishing their own future leverage. She has proposed that workers retain a share of revenue generated by systems built on their knowledge. Korinek's warning may be the most far-reaching: if the most ambitious projections about AI capability prove accurate, the real crisis will not be about how to compensate the trainers. It will be about what an entire society does when the knowledge economy no longer needs most of its knowledge workers.
Brendan Foody was twenty-two when Forbes named him one of San Francisco's youngest billionaires. Now his company, Mercor, is rewriting what it means to work in the age of artificial intelligence—and doing it by paying some of the smartest people in the world to teach machines how to do their jobs.
Tens of thousands of qualified professionals are collaborating with Mercor on a task that would have seemed absurd just a few years ago: training advanced AI systems to master the very disciplines these experts spent years learning. The company, founded less than three years ago, manages teams that instruct models built by OpenAI and Anthropic across fields ranging from consulting to philosophy. It is a phenomenon that is quietly reshaping what skilled work looks like and raising urgent questions about what happens to knowledge workers when the machines they train become better than they are.
The economics are striking. Mercor pays roughly two million dollars daily to approximately thirty thousand specialists. The average hourly rate exceeds ninety-five dollars, but in fields where demand is particularly acute—radiology, for instance—trainers can earn as much as three hundred seventy-five dollars per hour. These are not permanent positions. Projects typically last several weeks with no guarantee of future work. Yet they attract people seeking both flexibility and income that far exceeds what traditional employment offers. Beneath the appeal, though, runs a current of unease: the knowledge being transferred to machines might eventually render the knowledge workers themselves unnecessary.
Foody frames this differently. He describes Mercor as building a new "category of work" centered on training what he calls "AI agents." In his vision, the future workplace looks like this: everyone manages dozens or hundreds of these agents, spending their days interacting with and refining these systems to accomplish economically valuable tasks. He acknowledges that jobs will disappear. But he argues that entirely new categories of work will emerge to replace them, and Mercor positions itself as helping workers navigate this transformation.
Not everyone shares this optimism. Sadiq Khan, the mayor of London, recently warned of "mass unemployment" as artificial intelligence advances through finance, professional services, and creative industries. Anton Korinek, who directs the Transformative AI Economy initiative at the University of Virginia, told the Financial Times that these technologies could realistically substitute for "many of the functions" that define knowledge work. His observation cuts deeper: "The teacher is always eventually replaced by the student." The question is not whether displacement will happen, but how quickly and at what scale.
Among those working for Mercor, responses vary. An eighteen-year-old contractor cited the high pay, flexibility, and stimulating environment. Another reasoned that AI will "assist, not replace," and preferred to seize the opportunity now rather than watch others do it later. Amjad Hamza, a permanent Mercor employee, acknowledged social concerns but dismissed them as external to the actual experience: he sees a historical pattern where people work less while accomplishing more. Jay Katoch, a consultant, described his work in concrete terms: he presents models with scenarios a company might face after a disaster—managing damage control, handling stakeholders—and evaluates how the AI responds. The goal is to stress-test the systems and refine them.
The scale of this labor force is notable. Research commissioned by Scale AI and conducted by Oxford Economics found that forty-one percent of AI trainers hold master's degrees or doctorates. Ninety-four percent combine this work with other jobs or studies. The U.S. data annotation industry contributed five point seven billion dollars to GDP in 2024; projections suggest it will reach nineteen point two billion by 2030. Yet Zoe Cullen, a labor economist at Harvard Business School, identifies a structural problem: these trainers have no protections against the very models they are teaching. She proposes that workers retain a share of the revenue generated by systems built on their expertise. "If what you teach the model is your specialty," she argues, "by definition you're reducing your own bargaining power." Korinek's final warning is perhaps the most sobering: if the boldest projections about AI capability prove accurate, the real crisis will not be about compensating the trainers. It will be about what society does with everyone else.
Notable Quotes
The teacher is always eventually replaced by the student.— Anton Korinek, director of Transformative AI Economy initiative, University of Virginia
If what you teach the model is your specialty, by definition you're reducing your own bargaining power.— Zoe Cullen, labor economist, Harvard Business School
The Hearth Conversation Another angle on the story
Why would someone with deep expertise agree to train the system that might replace them? What's the actual draw?
The money is real—three hundred seventy-five dollars an hour in some fields. But it's also that these are temporary projects, not careers. People can do this for a few weeks, earn what they'd make in months elsewhere, then move on. There's no illusion of permanence.
So it's rational self-interest in the short term, even if the long term is uncertain?
Exactly. And some of them genuinely believe the technology will augment rather than replace. They see themselves as shaping something powerful, being on the frontier. That's its own kind of compensation.
But the economist Cullen makes a point about power. If you teach a model your specialty, aren't you literally eroding your own value?
She's right. The trainers have no stake in what their knowledge creates. They're paid once, then the model keeps generating value. It's a one-way transfer of expertise.
Does Foody's vision of everyone managing AI agents actually solve that problem?
It shifts it. Instead of being replaced, you become a manager of replacements. But that assumes new work emerges at the same scale and pays as well. Korinek's skepticism—that the teacher gets replaced by the student—suggests that's not guaranteed.
What's the timeline here? When does this actually matter?
The data annotation industry is already worth five point seven billion dollars. By 2030, it could be nearly twenty billion. The displacement isn't theoretical anymore. It's happening now, just not uniformly. Some sectors will feel it sooner than others.