AI Labs Pay Lawyers $400 an Hour to Train Models

 | 

Highlights

Mercor, a marketplace that places professionals with artificial intelligence labs, describes the job as applying professional judgment to how a model reasons and responds.

The labor that trains AI was long seen as the digital equivalent of blue-collar work. It now draws white-collar specialists who command much higher pay. Mercor told Bloomberg its workers average $110 an hour, though it said it did not have data on what share of tasks pay that rate.

Experts Write the Problems and Grade the Answers

Unlike the workers who record physical movements to teach robots how to perform tasks, these professionals are producing examples of intellectual work. They might solve a complex engineering problem, identify errors in a legal analysis or assess whether an AI-generated medical explanation is accurate. Their output becomes training material or evaluation data that developers can use to improve and test models.

The work looks less like labeling and more like setting an exam. Mercor’s legal listing asks experts to write a problem from a matter they handled, with the facts, forum and governing law, plus the reference answer a model’s attempt is graded against. They also grade model-drafted work, such as whether a privilege call is right or a citation supports the point it’s offered for. The listing pays $60 to $150 an hour.

Jack Weis, a Princeton chemical engineering Ph.D. who has applied to more than 2,300 jobs since graduating in 2024, took a $70-an-hour Mercor gig in January, The Wall Street Journal reported. He logged into a program called the Studio, wrote chemical engineering questions for a language model, then evaluated its answers and tried to improve them. Some prompts were simple, such as asking for an explanation at a sixth-grade level. Others involved technical specifications. The software recorded his screen, and he worked anywhere from four to 20 hours a week.

The expert pool keeps growing. Prolific, a British data-work platform, told Bloomberg experts make up 30% of its participants but didn’t disclose how many of its tasks they complete. LinkedIn ranked data annotator the fourth-fastest-growing U.S. job on its 2026 list.

Whether the work improves models is a research question. Researchers at Scale AI, a rival data vendor, found that checklist-style rubrics built from reference answers written by experts and stronger models produced significantly better models than rubrics generated by AI alone.

Handshake’s Top Rate Is $200 an Hour, and the Work Can Dry Up

Handshake, the student career network, runs a similar program. Its AI Fellowship signs students, graduates and professionals to project-based work training language models, according to its help center. Fellows are paid by Handshake, not by the AI labs. Pay is set in each project’s terms, either hourly or per task, and hourly projects pay for time spent revising and resubmitting work, its support site says.

Handshake’s physician posting lists up to $200 an hour depending on the project and asks for a medical license and direct patient care experience. A lawyer posting seeks a law degree, a license and practice experience at $90 to $110 an hour. Fellows in both roles design prompts and review model responses for accuracy and sound reasoning.

The work does not always last, as Weis found. The Mercor client, which the Journal said was never identified, gave him less work over time. Days without assignments became weeks, and in mid-March Mercor cut his contract. He learned of it from a Slack message sent to a group of 810 people.

Mercor Faces Breach Lawsuits as Its Valuation Climbs

The business model carries its own exposure. Mercor was hit with at least seven class action lawsuits in April after a data breach, PYMNTS reported. The suits say the breach exposed contractors’ interview recordings, facial biometric data and screenshots of their computers. Plaintiffs also allege Mercor monitored those computers, shared the data with clients and used recorded interviews to train AI models. Mercor has said it disputes what it called “speculative claims” in the suits. After its investigation, Mercor said in July that only a “limited subset” of experts had sensitive information leaked, Staffing Industry Analysts reported.

The breach hasn’t slowed the company’s dealmaking. Mercor announced July 9 it was buying Deeptune, a startup that builds simulated software environments for training AI agents. In August, Nvidia discussed joining a funding round that would value Mercor at $20 billion, double its valuation from October 2025, PYMNTS reported.

For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.