Lawyers training their own replacement: what it reveals about the real limits of legal AI

By El HuffPost · July 5, 2026.

Mercor, a company dedicated to recruiting experts from various fields to train artificial intelligence models, is paying between $100 and $200 an hour to retired judges and practicing lawyers to design deliberately complex legal scenarios and pinpoint where the AI's legal reasoning fails. The exercises range from simulating a corporate merger riddled with conflicts to building a complete fictional video game —with logos, licensing contracts and promotional posters— so the model can detect intellectual property infringements introduced on purpose. The stated goal is simple: if AI already codes fluently, let it also learn to reason like a lawyer, a far more slippery terrain than code.

The article, based on reporting from Business Insider, includes revealing testimonies. Jessica Crutcher, a U.S. attorney, explains that she takes on this side work to "stay relevant" and keep track of where her profession is heading. Charley Kelsey, who specializes in entertainment law, argues that AI will still be unable to reassure a distressed client, read the room during a settlement negotiation, or argue a case before a judge with the persuasion the courtroom demands. In other words: the very professionals feeding the machine are the ones best positioned to identify its limits.

There is something genuinely interesting about this business model, and it's not just the curious anecdote. It fits into a broader industry trend: major labs can no longer improve their frontier models with more internet data alone, so they turn to domain experts (doctors, engineers, now lawyers) to generate extremely high-quality synthetic training data, with intentional errors and edge cases that web scraping would never produce. It's the same pattern already seen in coding with benchmarks like SWE-bench: when a task is truly complex, training it well requires a human with judgment to mark what counts as sound reasoning versus a hallucination dressed up as a legal ruling.

The labor paradox is obvious and deserves to be stated plainly: these professionals are, in practice, being paid to meticulously document the instruction manual for their own partial replacement. It's the most literal face of the hard transition we've been noting in the legal sector: the base of the pyramid —routine contract drafting, document review, standard case-law research— is most exposed to automation, while expert judgment, emotional reading of a courtroom, and the trust relationship with a client hold firm. What's striking is that there's no need to speculate here about what will survive: it is the trainers themselves who, exercise after exercise, are mapping that frontier in real time.

Our reading is that this phenomenon, rather than a mere anecdote, works as a reliable thermometer of the real state of legal AI, far more honest than any product demo. If AI companies need to pay substantial sums to judges to invent traps the model can't solve, it's because high-level legal reasoning —one that weighs intent, cultural context and procedural strategy— remains a genuine bottleneck, not a data-scaling problem. In the short term, this doesn't prevent staff cuts in large law firms' administrative ranks, nor does it ease the pressure on junior associates whose routine tasks software already handles. But in the medium term, it points to a future where the lawyer who survives and thrives is precisely the one being paid by the hour today to teach the machine, with method and patience, everything it still doesn't know how to do: there is an almost optimistic irony in the fact that the automation process itself is financing and, however temporarily, dignifying the human expertise it aims to replace.