The Professionals Training AI to Replace Them — and What the Data Workforce Reveals

Laid-off professionals with advanced degrees are training artificial intelligence systems that will likely automate the jobs they once held. The work comes through platforms like Mercor, Scale AI, and Surge AI — companies that connect highly educated contractors with AI labs needing human-labeled data.

The irony is not lost on the workers. A former screenwriter told The Verge: “I’m being handed a shovel and told to dig my own grave.”

What the work actually is

AI models improve when humans provide training signals. The specific tasks vary but fall into a few categories documented in a March 2026 investigation by The Verge, in collaboration with New York Magazine.

Workers write “rubrics” — criteria for evaluating whether a model’s response is good. They produce “golden outputs” — ideal answers that serve as ground truth. They generate “reasoning traces” — step-by-step explanations of how to reach a correct answer. And they create “stumpers,” which are prompts designed to expose what the model cannot yet do.

Some projects involve “world-building,” where teams of contractors role-play corporate scenarios to generate documents, emails, and decisions that test how AI performs in simulated professional environments.

The workers come from diverse backgrounds. The Verge’s investigation identified Supreme Court litigators, McKinsey principals, chemists demonstrating poker and singing for voice models, a playwright earning $10,000 per month, a graphic designer who lost 85 percent of her traditional work to AI, and a Texan with a master’s in divinity training emotional tone in voice synthesis.

The platforms

Mercor, founded in 2023 by Brendan Foody, Adarsh Hiremath, and Surya Midha (then 19), was valued at $10 billion as of late 2025. Its clients include OpenAI and Anthropic. The company claims approximately 30,000 professionals work on its platform weekly.

Scale AI, founded in 2016 by Alexandr Wang and Lucy Guo, reports over 700,000 workers with master’s degrees, Ph.D.s, or college educations. In June 2025, Meta acquired a 49 percent stake in Scale AI for $14.8 billion, prompting Wang to join Meta and Jason Droege to become CEO. The company serves clients including Google, Microsoft, and the U.S. military.

Surge AI operates a platform called Data Annotation Tech, advertising projects staffed by Supreme Court litigators and McKinsey principals. Like its competitors, it has faced legal challenges over worker treatment.

Handshake, traditionally a campus job board, launched an AI data initiative connecting job seekers with training roles. Micro1 and Alignerr offer similar services.

Working conditions

The economics of the work deteriorate over time. A former freelance journalist and content marketer described being hired by Mercor through an AI interviewer named “Melvin.” Her first project — writing prompts, responses, and checklists — was paused abruptly after two days. She quickly accepted a second gig evaluating chatbot conversations, starting at 6:30 p.m. on a Sunday. A third project involved rubrics and stumping models, but her pay was cut by $8 per hour.

A graphic designer in her fifties worked on a Meta project tagging Instagram Reels videos at $21 per hour. When the project shifted to what workers called “Project Nova,” her rate dropped to $16 per hour — a 24 percent cut. Slack channels were deleted and direct messaging was disabled, leaving thousands of workers without communication channels.

Workers face surveillance through software like “Insightful,” which tracks keystrokes and deducts pay for time deemed “unproductive.” One former TV producer deactivated the monitoring software to work off the clock, avoiding penalties.

Termination comes without warning. Surge AI workers describe the “dash of death” — logging in to find an empty dashboard and a broken support button, signaling deactivation. A worker with a master’s in divinity reported feeling “cut adrift” after this happened.

Mercor spokesperson Heidi Hagberg told The Verge that “the nature of this is project based contract work, meaning it can extend, pause, or end at any time.”

Why the work is intermittent

The instability stems from how AI development works. Labs request specific batches of training data, pause projects to evaluate model performance, then change requirements as architectures evolve. A linguistics master’s holder who wrote rubrics told The Verge that her work dried up as models improved, leaving her unemployed for five months.

Strict non-disclosure agreements prevent workers from discussing project details, creating what Matthew McMullen, a strategy and operations executive, called structural silence. “The power is all on one side because they can’t talk about it,” he said. That silence prevents collective bargaining or workers leveraging shared experience.

The legal challenges

Three class-action lawsuits have been filed against Mercor in California within six months, alleging misclassification of workers as independent contractors despite the company exercising “extraordinary control” over how they work. One suit alleged a worker was forced to record sexually explicit scripts or face deactivation.

Scale AI has faced similar litigation. Lawsuits allege wage theft and psychological harm from exposure to disturbing content during data annotation. The company has settled some cases.

Glenn Danas of Clarkson law firm compared the platforms to Uber and Lyft, but noted a key difference: data workers are more replaceable and geographically dispersed than ride-share drivers, making traditional labor organizing harder.

What economists see

Daron Acemoglu, an MIT economics professor, compared the situation to pre-industrial weavers who lost autonomy when factory systems centralized production. He told The Verge that “we may also need unionlike organizations that exercise some sort of collective ownership and prevent any kind of simple divide-and-rule strategies by large companies to drive down data prices.”

The broader labor market shows signs of strain. Data from Handshake in August 2025 showed job postings declining more than 16 percent year-over-year while applications increased 26 percent — a widening gap between available work and seeking workers.

What the paradox reveals

The AI training workforce presents a structural paradox. The same expertise that makes a worker valuable for labeling data — legal reasoning, scientific knowledge, writing skill, creative judgment — is the expertise the model is being trained to replicate. Each rubric written, each golden output produced, each stumper crafted teaches the system what human professional judgment looks like.

The platforms benefit from this dynamic. They need skilled workers to produce high-quality training data, but their business model depends on models eventually performing those tasks autonomously. The more effective the training, the less human labor is needed — creating a workforce with an inherent expiration date.

Workers face a choice between participating in systems that may displace them and opting out without income. A former journalist who left Mercor for a coffee shop job described seeking “stability and human connection,” though she returned when a project resumed.

The legal battles over worker classification, the intermittent nature of the work, the surveillance infrastructure, and the structural silence imposed by confidentiality agreements all point to an industry still negotiating its relationship with the people whose expertise powers it.

Whether that negotiation produces institutional safeguards, collective bargaining mechanisms, or simply a rotating workforce of displaced professionals remains unresolved. The models keep improving. The workers keep labeling. And the gap between what the platforms need humans for today and what they will need them for tomorrow keeps narrowing.