At first, Laura thought the job advertisement was fake. The roles circulating in a WhatsApp group of her former classmates, promising $90 an hour for remote work in fields ranging from consulting to philosophy, seemed too good to be true.
When the business graduate joined Silicon Valley startup Mercor, the pay was real. But there was one big difference in this job from her usual management consulting projects. Her customers were not companies, but AI models from companies like OpenAI and Anthropic. Her role was to train the AI to perform the consulting work for which it was qualified.
“LLM training.” [large language models] “In order to do my job,” Laura said, the terms of her contract required her to use a pseudonym. In a tough labor market, she is interested in the job and feels that it is her preferred job above all others. But she worries about future unemployment. [my graduating class and I] I hadn’t really thought about it, but working with this kind of model gave me the feeling that this could be scary. . . As for [future] unemployment. “
Founded by three college friends, Mercor employs and runs teams that train AI to perform real-world tasks. It’s part of a growing subsector of AI training that includes Scale AI, in which Meta has a 49% stake, and Turing, which says it will “accelerate superintelligence” by using AI in simulated environments. In July, Surge AI, founded by Edwin Cheng, was reported to be in fundraising talks aiming for a $25 billion valuation.
Until recently, the freelance workforce training AI models was primarily low-wage, low-status workers from the Global South, responsible for labeling basic data and identifying harmful or traumatic materials. They are now increasingly made up of highly qualified contractors who are paid by the hour. They guide and evaluate the output of the AI, often training the model on how to perform advanced and economically valuable knowledge work. Mercor’s current roles on the platform range from journalism to real estate, but also include beauty therapy and social work.
Less than three years after its founding, Mercor raised funding at a valuation of $10 billion. It currently pays about $2 million a day to about 30,000 professionals. CEO Brendan Foudy, who became one of the youngest billionaires in San Francisco history at age 22, according to Forbes, said Melkor is creating a new “category of work” in training AI agents. Over the next few years, he says, more and more humans will need to fine-tune and advance AI and train it to do many of the tasks we currently do.
Foody paints a rosy picture of a future far different from the clunky, inconsistent, and error-prone AI tools that are familiar to many workers trying to understand new technology today. He says AI will take over day-to-day activities and act as an assistant, allowing people to move on to higher-level tasks where they may be able to do “incredible things that wouldn’t be possible otherwise.” In the meantime, the company says it is supporting workers by creating these new jobs and giving them exposure to advances in AI.
But to some observers, companies like Melkor seem like a short-sighted gamble, with skilled workers taking well-paying temporary jobs that train them in skills to replace them.
Already, there are growing concerns about the impact of AI on employment. London Mayor Sadiq Khan warned last week that AI risks creating “mass unemployment” in the capital unless steps are taken to protect white-collar jobs in sectors such as finance, professional services and the creative industries.
Anton Korinek, director of the Innovative AI Economics Initiative at the University of Virginia, said there is still a lot of uncertainty about whether AI will be good enough to significantly replace humans. But the current trajectory suggests that “many functions” of knowledge work could be performed in ways that “could be highly disruptive to the labor market,” such as taking over some jobs. Training AI still occurs within a “human-master, AI-apprentice model,” he says. “But at some level, the teacher is always replaced by the student.” [and] It remains to be seen how powerful AI will become. ”
At Mercor, project teams, ranging from a few contractors to hundreds, ask the AI questions, critique its responses, and provide step-by-step instructions on how to approach a problem.
One of them, Jay Katach, spends 40 to 80 hours a week working for Mercor. After decades in consulting, he says his work with AI is fulfilling. “I give [the AI] problem, [such as] How can a company like Boeing address the post-accident challenges of damage control, business impact, and stakeholder engagement? . . See how the AI responds,” he says. [the models] And to fix them. ”
Melkor says its average hourly wage is more than $95, but the most popular jobs, such as radiologists, can earn up to $375 an hour. Projects can span several weeks and continuous work is not guaranteed.
Zoe Cullen, a labor economist at Harvard Business School, said the short-term, gig nature of the work means workers are not protected if they help create a model that ultimately threatens their jobs.
She suggests that AI trainers could potentially keep a portion of the revenue generated by models that use their skills and knowledge. “Workers can participate in and benefit from increased productivity. They can also have collective agreements for data trainers,” she says. “If your core expertise is telling a model what to do, by definition you are reducing your workforce.”
Foody admits there is “absolutely”. [going to] “All kinds of new jobs will also emerge, so we want to help people overcome that displacement and help humanity and society achieve much more.”
Over time, he says, AI training will become a bigger part of more jobs. He sees a future where humans do things only once and AI creates rubrics to perform the same tasks over and over again. “We’re going to move into a world where everyone manages dozens or hundreds of agents, … interacts with those agents on a daily basis, and trains them to do economically valuable work.”
Many of Mercor’s contractors, anonymously at least privately, shared some of Foody’s confidence in the benefits of AI.
A study of the AI training industry conducted by Oxford Economics and commissioned by Scale AI found that most employees in the sector are highly educated, with 41% holding a master’s degree or PhD, and 94% also doing other types of work or study.
The US data annotation industry was found to contribute $5.7 billion to the US GDP in 2024, increasing to $19.2 billion in 2030.
“I and the people who work on these products have the experience of not losing their jobs, and I believe that those who don’t have that opportunity will,” said one 18-year-old contractor, praising the high pay, flexibility, and stimulating work environment.
Another contractor said he enjoys being on the “front line” of AI development. AI “supports, not replaces” and “even if I’m not making money doing it, there are plenty of other people out there trying to do the same thing.”
“There are concerns socially, but not within the group that is doing it,” Amjad Hamza, one of Melkor’s more than 350 full-time employees, said of the departures. “I’m just taking a very long view. The pattern in history is that people work less, but they can do more with that time.”
Sandeep Peechu, a managing partner at Felicis who led Melkor’s $10 billion funding round, says the company has become “very good at identifying the experts needed to train these models.”
He sees this as a positive for the future of his job. Melkor’s contractors are “appeasing and teaching models” in fields with labor shortages, such as nursing and law, he said.
Korinek of the University of Virginia is more pessimistic. If the most ambitious predictions about AI’s capabilities come true, Melkor’s trainers won’t just be out of a job, he says.
“If technology were more innovative, [than conservative estimates] “The big question is not whether the particular workers who train those models are compensated fairly. In essence, the big question is what do we do for everyone,” he says.
