A college student says Handshake AI owes him about $11,000 in back pay for hundreds of hours of AI training work — and he is not alone. Nearly two dozen current and former contractors say the startup underpaid them by hundreds or thousands of dollars each, with disputes over the past six months totaling $64,000, according to the contractors themselves.

Joel Thomas, a junior who transferred to the University of Mary Washington in Virginia, says he was ultimately paid $8,000 for work he believed was worth $19,000. Handshake has denied his appeals and deactivated his account, citing irregular activity. The company declined to comment on his case, saying it does not discuss individual matters. What does this change for the gig workers training frontier AI models? On the evidence here, the advertised rate and the paid rate remain two different numbers.

How the pay gap happened

Handshake, founded in 2014 by three Michigan Tech students as a university career platform, pivoted into AI training in January 2025 and rebranded to add AI to its name. It pays gig workers to help labs such as Meta, Google, OpenAI, Nvidia and Anthropic train and evaluate large language models (AI systems that generate text) through tasks like writing expert answers, designing edge-case test questions and rating model responses. Former staffers say the majority of Handshake's revenue now comes from AI-training contracts with those five clients. CEO Garrett Lord said in August the AI unit had grown past $1 billion in gross annualized revenue.

A Business Insider investigation, which reviewed bank statements, pay stubs and hundreds of emails and Slack messages, found several causes of the payment disputes. In many cases, contractor pay did not match expectations because Handshake and its clients tracked time differently. Contractors saw dashboard estimates based on manually logged timer hours; the client ran a second, invisible timer. Handshake compared the two before paying, and it did not tell contractors what counted as billable time or how the client calculated it, former employees said.

Staffing shortfalls compounded the problem. On some marquee projects, one manager oversaw up to 1,000 contractors, former staffers said. Project managers used AI coding tools like Claude and Lovable to generate instructions that were sometimes contradictory. If a confused contractor completed a task that failed review, the contractor went unpaid — a structure that shifts the cost of unclear instructions onto the worker.

A Handshake spokesperson said: "Final pay reflects the work our systems verify. Where logged time doesn't match the work, we review it, and where we find inflated time we don't pay for it."

What does the company's own marketing say?

The disputes sit awkwardly beside Handshake's public pitch. CEO Garrett Lord has repeatedly promoted transparent, reliable pay as an advantage over rivals such as Scale AI and Mercor, and has said complaints about late payments elsewhere motivated Handshake's entry into the business.

"Most AI gigs = shifting rates, zero transparency," Lord wrote in a LinkedIn post last year. "Handshake AI = competitive pay, clear expectations, and real dialogue. That's what experts deserve."

In October, after Business Insider sent detailed questions, Handshake introduced platform changes it said would improve pay transparency for contractors, whom it calls "fellows": a new payments dashboard, more explanations for adjustments, an easier dispute process and a live timer for work on third-party platforms. The company says any contractor who believes a payment decision was wrong can raise it again, and that a dedicated team independent of the project reviews disputes "in full and can restore all or part of the amount."

"Our Fellows are the heart of Handshake AI," the company wrote in an October blog post. "As part of being a Fellow you should always know what you earned, how it was calculated and when it will arrive."

That promise of transparency, it should be noted, arrived only after a reporter's questions.

The case of the 428 unpaid tasks

Thomas entered Handshake's orbit through Liberty University in Lynchburg, Virginia, which directed students to the platform. The university's own website disclaims endorsement of content, jobs or employers on Handshake and says use of its products cannot guarantee a positive outcome; the school did not respond to requests for comment.

After low-paying spring projects at roughly $17 an hour or less, Thomas received August invitations for projects paying $50 per hour and $65 per task. He worked long days from his bedroom desk until midnight or 1 a.m., watching his dashboard earnings climb past $19,000. One project, Project Folio, studied ChatGPT usage patterns and asked him to upload past chat logs; another required evaluating ChatGPT responses.

The payments that arrived fell far short. Handshake told him the "awaiting payout" figure was only an estimate, and that for the week of August 17 to 23, "our project leads did not find any tasking that took place during this time frame." Yet a pay statement Thomas received for that same period showed he had completed 428 tasks for Project Novel — with no pay attached. A spokesperson said contractors are not paid for unapproved tasks and that "no tasking took place" is shorthand for "no verified work"; the company may also deny payment for time exceeding a project cap.

"It's hard to do because it's still a lot of money," Thomas said of staying positive.

Business Insider first reported in March that dozens of contractors had complained about denied payments, and that one contractor had won a $6,475 judgment in small-claims court. Handshake says more than 188,000 contractors have earned over $500 million on its platform. Thomas, for his part, remains open to training AI models — but has given up on collecting anything more from Handshake.