When arbitrator Jessica Crutcher logs off after a day of hearing disputes, she starts a graveyard shift training AI models.
She gets the work through Mercor, a company that hires subject-matter experts to train artificial intelligence systems. She often spends hours staring at the walls of her home office, dreaming up legal problems that machines can’t easily solve, at least not yet. One week, she’s inventing a lawsuit over oil and gas trading; the next, she’s writing an asylum case. She prompts the model and then reviews its responses.
Lawyers have been told AI is coming for some of their work. Before that can happen, though, many are getting paid to teach it.
Crutcher is one of a growing number of legal professionals helping train the very systems that millions of people increasingly rely on. AI training companies like Mercor and Micro1 have enlisted thousands of lawyers, retired judges, and paralegals to date. Many spend nights and weekends crafting impossible legal puzzles, grading chatbot answers, and showing frontier models how experienced lawyers think.
The work offers a front-row seat to one of the biggest technological shifts to hit the profession — and, for many, a chance to help shape the tools that may one day reshape their own jobs.
More than 1 billion people now use chatbots like ChatGPT and Claude every month to draft emails and answer questions about legal squabbles from divorce to landlord disputes. The companies behind those systems want their answers to be as accurate as possible. To get there, they’re paying lawyers to supply something the internet cannot: professional judgment.
For Crutcher, the work is less about the paycheck than staying close to a technology transforming her field.
Investors have poured billions into startups chasing an opportunity to do for lawyers what vibe-coding tools like Cursor and Lovable have done for programmers. Software is learning to review contracts, look up case law, and handle other routine work that, for decades, filled law firms’ coffers. Keeping up with AI, Crutcher said, is part of how she “remains relevant.”
That same instinct led Harrison Margolin to start training models through Micro1.
His legal career had already put him close to Silicon Valley, advising semiconductor companies and software businesses. But the white-shoe law firm where he worked prohibited lawyers from using AI on client matters, and many colleagues dismissed the technology as a fad.
Margolin saw something different. Their reluctance, he said, gave him “more time to get ahead.”
After spotting a recruiting post from Micro1 on LinkedIn, Margolin signed up. Soon, he was being paid to stress-test frontier models.
One type of assignment is “red teaming,” or creating legal scenarios difficult enough that the models would fail. His toughest challenge was a cross-border merger packed with conflicts. He drafted fake legal documents and a scoring rubric, then kept making the exercise harder.
Micro1 kept sending it back with the same note. The model wasn’t sufficiently stumped.
The exercise illustrates one of the biggest bottlenecks in legal AI.
In programming, models can learn from billions of lines of public source code and immediately verify whether code works. Legal reasoning is much harder to teach, said Bertie Vidgen, a researcher at Mercor. Much of the world’s legal knowledge sits behind paywalls or inside confidential agreements. Even public documents rarely explain why lawyers made particular decisions.
That’s where human experts come in.
More recently, Margolin has shifted to writing “golden responses” — ideal answers that models can learn from rather than trying to trick them into failing. Isabel Yishu Yang, a special projects lead at Micro1, describes the work as “turning lawyers into law professors.”
The pay is respectable, though hardly Big Law money. Mercor advertises rates between $100 and $200 an hour for legal experts. Micro1 says compensation ranges based on different levels of seniority and role type.
The lawyers Business Insider spoke with all said the money was secondary to understanding where the profession is heading.
Entertainment lawyer Charley Kelsey reached the same conclusion.
While working at Lionsgate, he wasn’t allowed to use AI because clients prohibited it. Rather than wait for the legal industry to embrace the technology, he left to start his own practice and began taking projects through Micro1.
One recent assignment involved creating an imaginary video game, complete with screenshots, logos, licensing agreements, and posters, then asking the model to untangle the copyright and trademark issues he slipped into the materials.
The work unexpectedly changed how he practiced law.
Kelsey said lawyers often jump from a client’s question to an answer built on instinct and experience. Training AI, he said, forced him to slow down and examine every step of that reasoning.
He recalled one client asking whether a taxidermied bear could be copyrighted.
His instinct was no.
But after working through the analysis the same way he teaches models to do, he discovered case law treating taxidermy as sculpture — and therefore potentially eligible for copyright protection.
Eventually, the side gig stopped being a side gig.
In June, Kelsey gave up practicing law full time and joined Micro1.
He doesn’t believe he’s automating lawyers out of existence. AI may draft contracts faster or summarize legal research, but it still can’t reassure anxious clients, read the room during settlement negotiations, or argue before a judge.
What artificial intelligence will squeeze, he argues, is the work that didn’t require much judgment in the first place.
