A new labor market is emerging in the AI economy—one where professionals are paid handsomely to train the systems that may eventually replace them. Mercor, a startup now valued at $10 billion, has hired more than 30,000 contractors in 2025 alone to help refine AI models for clients including OpenAI and Anthropic.
The pay scale reveals which expertise AI companies value most. Dermatologists command $250 per hour, investment bankers and legal experts earn $180-200, poets surprisingly fetch up to $150, while video editors start at $45. The work involves reviewing AI-generated content, rating outputs against professional standards, and writing captions that help models understand nuance.
The Paradox of Training Your Replacement
The professionals taking these contracts understand the irony. One laid-off automotive journalist with 20 years of experience now spends 20-30 hours weekly critiquing AI-written news articles—essentially teaching the machine to do his former job better. “You’re training your replacement,” he acknowledges, “but the bills don’t pay themselves.”
Mercor’s client roster reads like an AI industry who’s-who. The startup sources talent across a remarkably diverse spectrum: astronomers, psychologists, industrial engineers, filmmakers, comedians, legal experts, and venture capitalists. Each brings domain expertise that helps AI systems move beyond pattern matching toward genuine understanding.
The Surveillance Layer
There’s a twist that highlights AI companies’ anxiety about their own technology. Contractors work under time-tracking software specifically designed to prevent them from using AI to critique AI. The systems monitor for ChatGPT usage, screenshot activity, and copy-paste patterns. The implication is clear: human judgment remains the gold standard, even as these same humans train machines to approximate it.
Structural Implications
This labor market represents a new form of knowledge arbitrage. Professionals monetize expertise accumulated over careers, converting tacit knowledge into training data. The transaction is asymmetric—once transferred, that knowledge becomes part of a model that scales infinitely while the human contributor returns to hourly work.
Legal experts reviewing contracts report a concerning pattern: IP agreements around “existing and future work” remain ambiguous. Some worry they’re signing away more than training data. Yet economic necessity overwhelms legal caution when the alternative is unemployment.
Mercor’s $10 billion valuation suggests investors see enormous value in this human-AI interface layer. The company has become infrastructure—a marketplace connecting AI labs hungry for human feedback with professionals willing to provide it. As AI capabilities expand, the question isn’t whether this market will grow, but how long human trainers will remain necessary.
Source: WSJ, Mercor company data
