The Work of Helping A.I. Destroy Work

Every day, Mercor, a start-up that sells training data to artificial intelligence companies, pays 30,000 contractors more than $4 million to help make their jobs, and those of their colleagues, obsolete.

It’s gig work, but for professionals with rarefied skills. One recent Mercor posting offered $225 an hour for a voice actor able to maintain a customer service persona in fluent Hebrew. Another sought a Ph.D. physicist with a specialization in general relativity, astrophysics or cosmology. A third listing wanted a physician with more than three years of experience in the Rwandan primary care medical system.

Mercor and a handful of similar start-ups are the primary middlemen in a supply chain of “human data” that may power the next generation of A.I. As OpenAI, Anthropic and other major ventures compete to become the industry’s dominant platform, the market for premium data that has been vetted by experts is exploding.

No longer do the A.I. companies need armies of low-paid workers, often overseas, to do rote tasks like tag images of cars or transcribe audio. They need mathematicians to annotate proofs, lawyers to mark up briefs and professors to grade essays. That’s what Mercor and its rivals supply. To use the parlance of the industry, data labeling has moved up the “value chain,” and the start-ups that offer this service have become some of the fastest growing in Silicon Valley.

Mercor, which was founded in 2023, announced a funding round in October at a $10 billion valuation; on Thursday, Bloomberg reported that the company is now talking to investors about a deal at twice that level. Last year, Meta invested more than $14 billion in another data-training venture, Scale AI, in part to hire its chief executive. Handshake, a recruiting start-up that pivoted to data training only in 2025, says its annualized revenue rate crossed $1 billion in April, up from $550 million at the start of the year.

These data-training start-ups are exploiting a market opportunity: selling to well-capitalized labs a product for which there is now near-unlimited demand. But it’s a delicate moment. The training start-ups need ChatGPT, Claude and other A.I. models released by their clients to keep improving, to demonstrate that they are adding value; they also need the models to remain imperfect so that those clients keep coming back for more data.

It may turn out that once OpenAI, Anthropic and others have taught their models to perform a certain job, their need for more training data in that area could sharply decline. In this way, Mercor, Scale, Handshake and their peers are much like the elite freelancers they employ: making money today, but in danger of being dropped tomorrow.

The start-ups would love to diversify their business so that they are less reliant on just a few major clients. For now, one way they are seeking to maintain momentum is by pressing into a tricky new area. Rather than just capturing the work of individuals (the Hebrew-speaking voice actor, the Rwandan physician), they are trying to capture the output of entire companies. Brendan Foody, the 23-year-old chief executive of Mercor, told me that this was “the bottleneck to a frontier lab automating everything that people do.”

On Thursday, he announced the acquisition of Deeptune, a start-up that makes “training environments” with simulations of the software programs, like Slack and Salesforce, that many workers toggle between all day long to get their work done. The idea is to painstakingly create a mirror image of, say, an investment bank so that A.I. can observe every interaction.

“One person will role-play the customer, the others will role-play the investment bankers, and they’ll build this out corresponding to the exact distribution as closely as possible,” Mr. Foody said. In theory, A.I. can then answer a profound question: “Like, what do they do at Goldman Sachs?”

What happens next to the workers at Goldman goes without saying.

Why would people participate in feeding their own careers into the A.I. wood chipper? People sign up for data-training gigs for a variety of reasons. The main one is, of course, money. This work is not glamorous, but it is in demand, which is more than can be said about many jobs, especially ones in crumbling fields like academia. Though the labor is unpredictable and rates vary — sometimes people are paid only if a supervisor approves the output — the workers who cobble together enough shifts can generate meaningful income.

People might sign up because they have been laid off, or because they can’t find enough work in their field. They might do it because they’re eager to get “A.I.” on their résumé, or because they need extra cash in retirement, or even because they genuinely believe in the potential of A.I. and want to be part of making the models improve. Many people who contract for these companies understand that this is a short-term opportunity, a brief chance to train the models to automate jobs before they themselves are automated out of the job of training models.

For Amanda Brown, an assistant professor of biology at Tarleton State University in Texas, part-time gigs on platforms such as Mercor and Handshake seemed at first like a straightforward way to make extra cash last summer while she wasn’t teaching. Some tasks offered $60 an hour, which sounded appealing, and she was interested in learning more about the models, especially since her students use A.I. all the time.

The experience “went downhill pretty fast,” Dr. Brown said. She started getting pulled into mandatory virtual meetings and working until 2 a.m. to meet deadlines. She became frustrated with gigs that paid a flat rate but took far longer than anticipated. Her Handshake manager told her vaguely that her work wasn’t good enough, which she found “soul destroying.” She did the gigs on and off until December.

“It seems like easy money,” she said. “It really isn’t.”

Part of the challenge, Dr. Brown said, was that over only a few months, she noticed rapid improvements that made it trickier to find things the A.I. models didn’t already know, which made her shifts that much more difficult. This summer, she found a teaching gig.

I wanted to see what the data-training start-ups looked like, and Mercor invited me to its San Francisco office for a tour and an in-person interview with Mr. Foody. Then, New York magazine published a report headlined “The Laid-off Scientists and Lawyers Training AI to Steal Their Careers,” which painted a harrowing picture of the unstable, inconsistent, emotionally draining labor offered by Mercor. “I have never, ever been treated as badly as this,” one contractor told the magazine. Mercor canceled my visit. (The company later made Mr. Foody available for a short Zoom interview.)

A common complaint in the Discord channels and Reddit threads where contractors gather is that A.I. training work has dried up in recent months — that, in other words, once a set of experts teaches the models how to do something, their services are no longer needed in the same way. Some of the gig workers I spoke with told me that, in the past year, the work had gotten significantly more intricate. Sommer Wall, a lawyer who has contracted with Mercor, said she sometimes napped on her couch during 72-hour grinds to meet deadlines on the most intense projects.

Dissatisfied contractors — many of them sophisticated corporate operators — have brought legal action claiming that Handshake, Scale, Surge AI and other data-training start-ups are misclassifying, underpaying or otherwise exploiting them. After a data breach this spring, Mercor faced at least seven lawsuits from contractors who said their personal information had been exposed. (Most of the cases are ongoing.)

As contentious and even toxic as the work may be, many more people may soon be pushed into these freelance gigs if even some of the predictions about A.I. job loss come true. Confidence in the labor market has already collapsed, including among highly educated people: Just one-fifth of college-educated workers said now was a good time to find a quality job, according to recent Gallup polling.

Willing or not, specialized workers who entered their careers expecting top compensation and prestige may find themselves among the contractor masses. Mercor seeks people with “strong academic credentials (M.A.s, J.D.s, M.D.s, Ph.D.s, etc.) from top universities.” Handshake, perhaps sensitive to the egos of academics, calls its A.I.-labeling opportunities a “fellowship” and seeks, per one listing, “exceptional master’s and doctoral students, candidates and graduates.” Elon Musk’s lab, xAi, which dabbles in in-house training, has solicited writers who have sold 50,000 copies of a novel or published 10 short stories in elite magazines for a gig starting at $40 an hour. A recent listing from Surge called for people with a “top-tier consulting pedigree” to “capture partner-level reasoning.”

“If you want to navigate and not get replaced, you will need to embrace these skills and this knowledge to accentuate your job,” said Jonathan Stull, the chief operating officer of Handshake. He was sitting at a large wooden table in a conference room at his company’s headquarters in San Francisco this spring, arguing that most white-collar jobs will soon involve giving feedback to A.I. models. Taken to its logical conclusion, this is a vision of work in which even high-level employees systematically automate nearly every task they perform, until all that is left for them to do is train A.I.

Handshake is hoping to make billions of dollars hastening that future. In the conference room, Mr. Stull and the company’s head of research, Paco Guzmán, explained how information is transferred from the minds of professionals into the models. Handshake does not disclose which A.I. labs it counts as clients, but in recent months it has publicized collaborations with Google and OpenAI.

Broadly speaking, models develop in two ways: through pretraining and post-training. The former is what many people think of when they picture how A.I. learns — by sucking up the corpus of text and information on the internet. But even a full scrape of the web will not garner the sum of human knowledge. Quality information has become harder for A.I. to reach, as various companies have restricted access to their content. (The New York Times has sued OpenAI and Microsoft, claiming copyright infringement; the companies deny the claims.)

To keep improving their models — to make them more useful, more sophisticated, less prone to hallucination and mistakes — A.I. companies heavily refine what goes into them. That’s post-training, and it includes buying data from vendors like Handshake and its competitors.

At Handshake, a “fellow” might spend a shift putting together detailed samples of, and solutions to, problems that they might encounter in their area of expertise; or flag which of two responses to a given prompt is better. When a prompt doesn’t have an objectively correct answer — say, when the piece of work onscreen is not a math problem but an essay — evaluators can score responses according to a rubric. Does the essay have a clear structure? If so, it can be awarded a certain number of points. Does the essay use proper grammar? That could increase its score, too.

Before submitting training data to a client, Mr. Guzmán said, Handshake performs some quality control, checking data and using automated tools to assess whether the contractors might have plagiarized or used A.I. to complete their work. (Yes, trainers using A.I. to improve A.I. is forbidden.) Handshake then sends its clients a database with hundreds or thousands of values, containing the original prompt, the top answer and various metadata.

Sometimes, clients approach Handshake with something they need and it’s up to Handshake’s operations and research teams to figure out how to deliver it.

“How do you structure this?” Mr. Stull said. “How do you find the right people? How do we get them in? We have to pay them, we have to vet them, we have to trust them.”

And sometimes, Handshake approaches clients and persuades them that they need something — by demonstrating where their available models are falling short, for example.

During my visit, Handshake was working on what it called a “banker benchmark” — effectively, a standardized test to gauge whether an A.I. model could perform the tasks expected of a human working in finance. Mr. Guzmán projected a long, detailed document on a big screen. The benchmark included a role-playing scenario addressed to an A.I. model: “You’re a junior investment banker,” the prompt section began, laying out a situation in which a managing director is asking an underling for a client report. The next section showed an “ideal output,” and another section listed some 90 weighted criteria for evaluating whether the model solved the problems properly.

“We are verifying not just their answer but that, along the way, they actually made the right decisions,” Mr. Guzmán said. The model needs to show its work, not just spit out the solution. Human contractors, he noted, spent many hours meticulously putting together the content of the benchmark. In theory, an A.I. model that gets a high score is ready to take over a banker’s job.

The data-training start-ups see a lucrative opportunity in recreating workplaces in miniature: controlled environments in which their gig workers can evaluate and reproduce emails, memos and slide presentations in context. The information emerging from such a setup, the companies boast, will help shrink the gap between what A.I. models can accomplish and what office workers actually do from one minute to the next, as ideas and instructions flow between meetings, documents and applications.

Scale, for example, has said that “our environments replicate real-world workflows,” and that its contractors “curate artifacts that capture the complexity, ambiguity and edge cases of real professional work.”

Executives see the models’ shortcomings as a sign there’s more for them to do.

“I often use Claude Cowork, right?” said Edwin Chen, the founder of Surge. “And even though Claude Cowork is incredibly smart, oftentimes it doesn’t quite understand the nuance of Slack. It doesn’t quite understand this ambiguous question I have. It doesn’t quite understand where to go and find this Google document.”

According to the Bloomberg Billionaires Index, Mr. Chen’s stake in Surge — and his vision for what it could become — makes him the 258th-richest person in the world.

“I often think about us as essentially, like, the school for A.G.I.,” Mr. Chen said, referring to a prophesied level of A.I. that surpasses human intelligence. “A.I. comes to us, and A.I. learns to run the world.” He and his customers at the big A.I. labs, he said, are designing the curriculum.

Mr. Chen, who styles himself as something of a thinker in an industry of merchants, mused about whether “education will still be a thing” in the future, or whether people will just watch A.I. slop videos on their phones for entertainment all day. But he argues that his company is creating paying work, as well as helping to one day spawn A.I. collaborators that will not replace jobs but rather assist knowledge workers in getting more done.

Intellectuals, he predicts, will eventually embrace the mental challenge of training and improving A.I. — and consider it a novel career path that is “one of the most prestigious and impactful” in society.

Executives with a vested interest insist that demand for training will continue indefinitely, but others are not so sure.

“As A.I. systems get better, there is going to be less of a need for human teachers for the A.I.,” said Anton Korinek, an economist at the University of Virginia who just began a yearlong leave of absence to work at Anthropic. “Both the importance of that kind of post-training data overall, and the growth of companies specializing in that, may decline somewhat.”

As for the prediction that most white-collar employees will do some form of data training, Dr. Korinek said he thought it was too general. “It kind of sounds like saying, in the future, all white-collar workers will be professors,” he said. “Why would that be true?”

The concept of artificial intelligence implies an obviation of the human. Even its name suggests the idea that the machine does its thing, no human necessary. But, at least so far, a fleet of humans is playing an intensive, hands-on role in developing that intelligence.

“The machine is hiding the things that humans are doing,” said Margaret O’Mara, a historian of Silicon Valley at the University of Washington. She is careful not to overestimate either human or machine intelligence. Not only might the models run out of data to train on, but humans might run out of ideas to give them. This big knowledge transfer of human to machine “is happening at a moment where there is a diminishment of human expertise because of A.I.,” she added.

The industry could simply lose momentum before it remakes society. “It’s worth not just taking this as a foregone assumption that this is the way it’s going to play out,” Dr. O’Mara said, “because that’s how these companies are telling us it’s going.”