Their New Job Is Teaching AI to Replace Them

Next, it could be you.

Crossing Editor's Note

During the National Day holiday, we'll be sharing a series of articles worth reading. Today's piece of nonfiction departs somewhat from Crossing's usual fare. We've been tracking entrepreneurial opportunities and innovation amid the AI wave. This time, we want to turn our attention to the people who supply AI with professional expertise and train its models: how they work, and how they make a living amid constant change. To understand AI's transformation, we must also see the people caught within it. This article comes from The Verge and New York.

After AI Took Her Job, She Went to Train AI

The LinkedIn post looked like another recruiting scam, but Katya was desperate enough to click. After college, she'd worked as a freelance journalist, barely scraping by. She went to grad school, then pivoted to content marketing, hoping for something more stable. Instead, she found that AI had automated much of it.

A company called Crossing Hurdles promised copywriting work starting at $45 an hour.

Katya clicked through and was taken to a company called Mercor. The page asked her to turn on her camera for an AI interview with someone named Melvin. "This looked extremely sketchy," Katya said. She closed the tab.

But weeks later, still unemployed, she received a message inviting her to apply for work at Mercor. This time, she looked into the company. It turned out Mercor sold data used to train AI, and it was recruiting her to produce that data.

"My job was eliminated because of ChatGPT, and now someone was inviting me to train the model to do that job in the worst possible way," she said. The thought depressed her. But her finances were deteriorating, and she urgently needed a new place to live, so she turned on her camera and said "hello" to Melvin.

It was a strange but overall pleasant experience. Melvin appeared on Katya's laptop as a disembodied male voice. He seemed to have actually read her resume and asked specific questions about it.

Weeks later, Katya received a job offer email from Mercor. Like most workers in this story, she asked to use a pseudonym, fearing retaliation. The email said that if she accepted, she would need to sign a contract, pass a background check, and install monitoring software on her computer. She signed immediately.

She was added to a Slack channel that was clearly part of a project already underway. Hundreds of people were busy writing example questions that users might ask a chatbot. For each, they would draft an ideal response, then compile an exhaustive list of criteria defining what made that response ideal.

Each task took hours to complete. Afterward, the data would be sent to workers further down this digital assembly line for additional review.

No one told Katya whose AI she was training — managers referred only to "the client" — or what the project was for. But she enjoyed the work. Working with the model was interesting, and the pay was good. "It felt like having a real job," she said.

Two days after Katya started, the project was abruptly paused. A few days later, a supervisor appeared in the channel to announce that it had been canceled. "I thought I could plan around this. I was saving up for first and last month's rent on an apartment," Katya said. "Then I was back to square one. No warning, no security, nothing."

A few days later, Mercor sent another job offer. This time, the task seemed to involve evaluating conversations between chatbots and real users according to various criteria. Many of the conversations appeared to be users in Malaysia and Vietnam practicing English. The evaluations covered how well the chatbot followed instructions and whether its tone was appropriate.

The email said to sign the contract and attend a Zoom onboarding session in 45 minutes. It was 6:30 on a Sunday evening. Still shaken by how suddenly her last gig had vanished, she accepted and worked until she was too exhausted to stay awake.

Human Expertise, Bought in Bulk

Machine learning systems learn by finding patterns in vast amounts of data, but that data must first be organized, labeled, and produced by people. ChatGPT's fluent articulation depends on thousands of workers employed by companies like Scale AI and Surge AI. They write the things a helpful chat assistant would say and score its best responses.

A little over a year ago, the industry began growing increasingly worried that the technology's progress would plateau. Chatbots trained on such ratings had become adept at sounding intelligent, yet remained too unreliable to be genuinely useful.

Software engineering was an exception. In this domain, models could automatically verify whether code worked — whether it compiled, whether it printed HELLO WORLD — and thus could truly master the skill through trial and error.

The problem is that few other human activities offer such clear-cut feedback. There is no objective test for whether a financial analysis or an ad campaign is "good." AI companies did not retreat in the face of this. Instead, they set out to create such tests, paying billions of dollars in total to professionals of all kinds to write rigorous, comprehensive standards defining what it means to do a job well.

Mercor, which Katya had stumbled upon, was founded in 2023 by three Bay Area teenagers — Brendan Foody, Adarsh Hiremath, and Surya Midha — who were all 19 at the time.

It began as a recruiting platform that used AI interviews to match tech companies with overseas engineers. Then AI developers began flooding in with requests: Could you help find professionals to produce training data? The demand was so great that the company pivoted.

Last year, Mercor reached a valuation of $10 billion, making its three founders the world's youngest self-made billionaires. OpenAI was a client. So was Anthropic.

These data companies all emphasize the impressive credentials of the experts in their ranks. Mercor says about 30,000 professionals work on its platform each week; Scale AI claims more than 700,000 "master's, PhD, and college graduates."

Surge AI advertises experts including litigators who have argued before the Supreme Court, McKinsey & Company senior project leaders, and platinum-certified musicians. The companies are recruiting people with legal, financial, and programming experience — all fields AI is rapidly penetrating.

But they are also hiring for data production in virtually every occupation you can think of. Job listings include chefs, management consultants, wildlife conservation scientists, archivists, private investigators, police sergeants, journalists, teachers, and rental counter clerks.

One recent ad sought an expert in "humor for early-to-mid teenage years in North America." Among the requirements: ability to "explain humor in clear, logical language, including references to North American slang, trends, and social norms." As one industry veteran put it, this is the largest acquisition of human expertise ever undertaken.

These companies have a vast pool of recruits in the growing number of highly educated but underemployed people. Hiring activity is at its lowest level in decades, outside of the 2008 financial crisis and the pandemic.

Last August, Handshake, a job platform for early-career professionals, found that job postings on its site had dropped more than 16% year-over-year, while applications per job had risen 26%.

Meanwhile, Handshake launched a program last year connecting job seekers with positions producing AI training data. In announcing the program, the company wrote: "As AI reshapes the future of work, we have a responsibility to rethink, educate, and help the people in our network prepare for and plan career paths that engage with the AI economy."

People predict that general intelligent systems will replace vast amounts of human cognitive labor. But what AI labs are actually doing is paying for data to automate tasks one by one. There is an inherent tension between these two visions.

They correspond to two different futures: one of sudden mass unemployment; the other less visible but potentially equally disruptive — more and more people finding that their new job is teaching AI to do their old one.

The first to travel this path were software engineers, graphic designers, writers, and professionals in other fields where new training methods are showing results. They have found themselves in an absurd position: scrambling for unstable gig work that involves simulating and performing the careers they once aspired to.

Showing AI How Humans Work

I interviewed more than 30 workers, each situated at some point along a vast and ever-expanding data supply chain.

Some write lists defining what a chatbot's good response should satisfy. These lists are typically called "rubrics." Others score these rubrics. Still others use rubrics to judge chatbot responses, while some use rubrics to write what's commonly called "golden output" — the ideal chatbot response.

Others are asked to explain, in the voice of a chatbot talking through its reasoning, every step by which they arrived at the golden output. This produces what's called a "reasoning trace," which the AI can later consult when encountering similar tasks in the real world.

Sometimes, labs want rubrics only for problems the model currently cannot solve. This means companies like Mercor ask workers to write "stumpers" — requests designed to make the model fail. "It sounds easy, but it's really, really hard," said one worker who was trying to stump models by asking them to build inventory management dashboards.

Models fail in counterintuitive ways. They might solve advanced physics exam problems, yet if you ask about public transit routes, they'll suggest transferring between two lines that don't intersect. Finding these weaknesses takes time and creativity.

Another type of project assembles lawyers, HR managers, teachers, consultants, or banking professionals to do what Mercor calls "world-building." Training materials state: "You and your team will simulate the work of a real team in your profession through role-play."

Each team is given dedicated email inboxes, calendars, and chat applications, and asked to produce a hundred or more documents around some corporate business activity — for example, a fictional mining company analyzing whether to enter the data center business.

One worker recalled spending 16 hours a day for several days straight producing these fictional documents. Then the presentations, meeting minutes, and financial projections would be handed to another team, which used the materials to try to stump models operating in this simulated corporate environment. After succeeding, they would write more detailed new scoring rubrics, standard answers, and so on.

Workers could only guess who the client was, or how many people were involved in the project. Based on team names like "Management Consulting World No. 133," there might be hundreds of people, or thousands.

Some were hired to assess image models' ability to follow prompts. Others described video clips in extreme detail, presumably to train video models. To improve AI's voice conversation abilities, demand for voice actors also surged. One job posting said they might be asked to record speech that was "authentic and emotionally resonant."

"I just tell people I'm an AI trainer — it sounds more professional than what I'm actually doing," said one aspiring screenwriter. He had been asked to record himself pretending to ask a chatbot for a fitness plan, while pots and pans clattered in the kitchen. Another time, he was asked to record himself giving financial advice over the phone to wave after wave of people; he guessed they were also other workers.

This audio would then likely be broken apart and handed to people like Ernest. He used to make a living through online tutoring, until the company that employed him replaced him with a chatbot. When we spoke, he was listening to random conversations of roughly one minute each at 0.1x speed, marking the milliseconds when speakers started and stopped talking.

In many clips, a person was talking to a chatbot, occasionally inserting "ums" or "got its." So he guessed he was helping AI improve its ability to sound natural and fluid in conversation — though he didn't actually know.

As is customary in this industry, projects are referred to only by code names, and clients are forever just "the client." The entire system is designed so workers know little about the supply chain they inhabit. Even if they figure out who the client is, contracts forbid them from telling anyone, including their own colleagues.

According to one Mercor agreement, they also cannot describe the specific work, only say vaguely that they are "providing expertise in such-and-such field to help a top AI lab improve its models."

Workers live in fear of accidentally violating confidentiality agreements and getting fired. So when discussing work in public forums, they add another layer of code names to projects that already use code names. For example, calling a project named "Raven" by the name "Poe."

He started working for free

"Someone handed me a shovel and told me to dig my own grave." Katya's second project at Mercor was far more stressful. There was less work available, and it came and went. Managers would post messages in Slack channels saying new tasks would arrive in half an hour. She said afterward, "Everyone in Slack would drop what they were doing and swarm like piranhas," racing to speed through tasks as they watched the progress bar showing remaining tasks tick down to zero.

Then they would return to Slack, politely pleading with supervisors for more work, more hours. They would mention their child's birthday, the rent they needed to pay, or tell anyone who might be listening that they were available anytime if there was more work.

Soon, Katya too became someone who would drop everything at the sound of a Slack notification. "Sometimes I'd be in the bathroom or eating dinner, the Slack notification would come, and I'd say: 'Sorry, I have to work now.'"

That project ended quickly, and another came right after. It was almost identical to the first project she had enjoyed, but now, in addition to writing scoring rubrics, she had to stump the model too — completing harder tasks in the same amount of time. Her hourly pay was also $8 less.

This was common at Mercor. Nearly every worker I interviewed said that as projects progressed, demands increased, allotted time shrank, and pay dropped. Those who couldn't meet the new requirements were "removed from the project," replaced by newly hired people.

Chris joined Mercor last year, after months of struggling to find film and television work. While many merely suspected they had become victims of automation, he was certain he had.

He used to have steady work writing episode content for an unscripted TV show: conducting pre-interviews, designing scenarios, writing reality-show versions of scripts. But by late 2024, he was told the show would run with "minimal staffing" and no longer needed his work.

Later he discovered the company was using ChatGPT to draft content for new episodes. So when he was invited in October to write an entire science fiction script for a major AI company, despite the bleak implications, he said yes.

Since then, he had taken one gig after another. "This is my only source of income right now," he said. "I know award-winning producers and directors who don't go around saying they do this kind of work, but this is how they're supporting their families."

Like Katya, his first few jobs at Mercor were fairly pleasant and paid well. But soon, every evening at 6 p.m., motivational messages with fist-bump emojis would appear on Slack: "Let's go, team, we got this." Then work would suddenly stop, followed by months of silence.

"You're just waiting for the starting gun to go off — it could happen anytime," Chris said. Then the rush would begin again. As deadlines approached, increasingly frantic managers began threatening workers: complete tasks fast enough or be removed from the project.

His work time was tracked to the second by software called Insightful, which monitored everything he did on his computer. Time the software judged as "unproductive" could be deducted from his pay; if he went several minutes without typing, a popup would ask whether he had been working continuously.

Sometimes Chris would see someone in Slack say that a particularly tricky task had exceeded the allotted time, hoping it wouldn't be a problem; the next day, that person would be gone.

Growing increasingly afraid he too would be removed from the project, he began working without logging hours: turning off Insightful while reading task instructions, so he would appear faster on the clock. If he exceeded the allotted time, he would turn off the timer and keep working for free.

The company said such software was necessary for accurately tracking hours and preventing workers from cheating. The "cheating" in question meant using AI — something all data companies strictly prohibit.

These companies sell verified human expertise as a reliable ground truth. And when AI is trained on AI-generated data, it gradually degrades. Researchers call this phenomenon "model collapse." Data company employees say filtering out low-quality AI-generated content is a constant battle.

For workers, the greater the pressure, the stronger the temptation of AI. The retail expert who tried to stump models with analytics dashboards, after task time limits dropped from 8 hours to 5 hours to 3.5 hours, turned off Insightful and sought outside help.

"Honestly, I opened Copilot and ChatGPT, put in the prompt and asked: 'What should I change to make you unable to answer?'" Then she asked another chatbot whether the prompt sounded AI-generated; if it did, she asked it to make it sound more human-written.

"The psychological impact on people is really terrible," said Mimi. She was a screenwriter who had worked on multiple streaming series, and had been training AI for Mercor for several months. She had learned about Mercor from a fellow screenwriter who posted the company's job link in a Facebook group for the Writers Guild of America.

Like many people doing this kind of work, Mimi was conflicted. "An Emmy-winning documentary director messaged me saying: 'Someone handed me a shovel and told me to dig my own grave.' That's how everyone feels," she said. But as a single mother, she needed the money.

At first, she was grateful to have the work. Then projects paused, restarted, paused again. For five straight weeks she was told a project was about to launch. When it finally started, requirements had increased, estimated completion times shortened, and she had to race to keep up under Insightful's surveillance.

She thought someone in Slack had put it well: everyone was like fish living in a tank, waiting for their human owner to drop in food — only the fish fast enough to swim to the top got to eat.

"Last night I was so fucking stressed. My kid came home, it was already 7 p.m., I got a message: 'Tasks are out!' So I started working, just trying to rack up as many hours as I could before bed," Mimi said, her voice breaking.

"I have no time to spend with my kid. There was a moment when he couldn't find something he needed for school, and I just started yelling at him. This job is turning me into a fucking monster."

What especially disturbed her was the surveillance: "Someone can measure your time this way, strip away all the minutiae of what you need as a human being, because those things don't generate profit. You can't bill for going to the bathroom, because that's not work time. You can't bill for making coffee. You can't bill for stretching because your back hurts."

"This is why unions were formed — to give people guaranteed hours, lunch breaks, vacation, sick pay. This is pushing the gig economy to the extreme."

This worried her more than AI itself: AI was bringing the unstable platform labor that had already transformed the taxi and food delivery industries into knowledge work. Meanwhile, she watched with horror as her colleagues cheered with desperate gratitude for work notifications posted at 7 p.m.

"How long are these tasks expected to last?" one worker asked in Slack.

"I want to know too, I want to know if I can sleep."

When no one answered for a long time, they began exchanging tips on how to stay awake.

Hourly pay suddenly drops 24%

"Nobody knows what's happening, everyone is confused." When Mercor began hiring aggressively last year, it positioned itself as a more worker-friendly option than previous platforms. Mercor CEO Foody criticized competitor Scale AI on a podcast, saying: "Having great people and treating them really well is the most important thing in this market."

Workers who joined during that period did say they were treated well: pay was higher than elsewhere, and they were managed not by the opaque algorithms common in the industry but by actual human supervisors they could ask questions.

But people who had worked as managers at data companies said this was how these companies always started: promising better treatment to lure workers away from existing platforms.

Then they began competing for eight-figure-dollar contracts. These contracts were issued by the half-dozen or so AI companies that purchased data in bulk. In this competition, working conditions deteriorated.

At Mercor, the situation had an added layer of complexity: the management consisted mainly of twenty-somethings with little work experience, yet investors had given them hundreds of millions of dollars and demanded high-speed growth.

"My manager is 21, I don't care about that," said reality show producer Chris. "But they've never done work at this scale before."

"You reach out on Slack for guidance, you explain the situation maturely and clearly, and what you get back is a corgi rolling its eyes with the text 'use your judgment.' But what it actually means is: 'Use your judgment, and if you screw up, you're out.' You went to Harvard, graduated last year, and when you're managing a group of people — many of them seasoned professionals — your guidance is a meme?"

Lawyers, designers, producers, writers, scientists — everyone complained about inexperienced managers. They issued contradictory instructions, demanded long hours, or imposed mandatory Zoom meetings on supposedly flexible arrangements. They also threatened that anyone who moved too slowly would be removed from projects.

This threat especially enraged mid-career professionals, who felt their twenty-something bosses barely understood the fields they were trying to automate.

"The founders are proud of '996,'" one lawyer said. The term originated in China to describe a 72-hour work week, associated with burnout and even suicide, yet Silicon Valley had adopted it as an aspirational work style.

"You have to be available around the clock. They start firing off messages at 6 a.m., and you'd better move fast, because everyone feels like if they don't, they'll be removed from the project and someone else will take their place."

"It's not just that team leads are young, project managers are young, senior project managers are young. Even the people above them, running entire projects — senior project managers — they're young too. I guess it trickles down from the top, since the founders themselves are young, right?" said Lindsay.

She was a graphic designer and illustrator in her fifties. Over the past year, 85% of her work had disappeared, which she attributed to advances in generative AI; after that, she came to Mercor.

Increasingly desperate for work, she scoured job boards. It seemed the only positions matching her expertise were ones asking her to help develop the very technology she believed had destroyed her career. "I swallowed my disgust and signed up," she said.

Initially, she did some graphic design data production work, then was invited to join a project for Meta: finding videos on Instagram Reels and labeling various content within them.

The work was tedious, and $21 an hour wasn't great, but Lindsay needed the money. After being added to the project Slack, she discovered roughly 5,000 others were in the same boat.

In early November, a Mercor representative announced that due to "scope changes," Lindsay's project was ending. Previously, workers had been told the project would continue through the end of the year. Lindsay and thousands of others found themselves removed from the company's Slack.

Soon, an email appeared in their inboxes inviting them to join a new project called Nova, at $16 an hour.

Thousands of workers flooded into the new Slack, only to discover the work was exactly the same, but the pay was 24% less. The Slack channels had been stripped down to just two; all others — including casual chat, support, and help channels — had vanished. The ability to direct message each other had been disabled too. Team leads were nowhere to be found.

With no one to turn to, workers resorted to posting requests and angry questions in the main channel.

"No one knew what was happening, everyone was confused," Lindsay said. "Messages in that channel were flying by so fast, it was complete chaos. 'Please, help me. What do I do? What should I do? Where do I go? Can I start working on tasks now? Do I have to redo all the tests I already did?'"

Someone emailed customer support for help, and for some reason, the email went out to over a thousand people on the project. Everyone seized the opportunity and started hitting "reply all" to express their confusion and anger.

"It was absolute mayhem," Lindsay said. "There's no other word for it."

Workers began posting complaints in Mercor's Reddit community, but the posts were quickly deleted by Mercor representatives who moderated the community. So two unofficial Mercor communities were created, where workers could freely express sentiments like: "This company is run by a bunch of kids, and they're going to pay for it eventually."

"It's really heartbreaking," Lindsay said. "For some people here, this really determines whether they can feed their families."

"I hate generative AI," she added. "I think AI should be used to cure cancer, to explore space, not in creative industries. But I have to pay rent. And then companies like Mercor pull something like this, treating you like a lab rat. I've worked for so many years, and never, never have I been treated this badly."

Kept Waiting

Intermittent work, extreme secrecy, and sudden termination are the norm across the data industry. On Data Annotation Tech, a work platform under Surge AI, workers are not only frequently terminated without cause — often no one even tells them they've been fired.

They simply log in one day to find their task dashboard empty. The phenomenon is so common that people have taken to calling it "dashboard death."

Last year, a Texas worker with a master's degree in theology was teaching voice models to respond with appropriate emotions to different questions. For example, a user saying their dog had died should receive a different tone than one asking the model to plan an itinerary.

One morning, he logged in to start work and found his dashboard empty. He scrolled to the bottom of the page to click the customer support button, only to find it had been deactivated. That's when he realized he'd been terminated.

His mind raced through possible reasons: Had he worked too much? Had his quality declined? He knew he would never get answers. "I felt cast adrift," he said.

Worried about paying bills and caring for his sick dog, he became depressed, then fearful. He thought of friends who were teachers, no longer able to get students to write on their own; and recent graduates with computer science degrees that now seemed worthless.

"This technology has led us to see everything as tools, as things to be used," he said. In his view, that included data workers like himself who were discarded. He resolved to become a pastor providing spiritual care, because he believed that whatever future AI brought, people would always need another person by their side.

The stop-and-start nature of this work isn't just a product of corporate culture — it also stems from the rhythm of AI development itself.

People across the industry have described this pattern: a model developer like OpenAI or Anthropic discovers its model is weak in chemistry, so it pays a data supplier like Mercor or Scale AI to find chemistry experts to produce data.

The chemistry experts work steadily until enough data is assembled to send back to the lab, then work pauses while the lab sees what effect the data has on the model. The lab may decide to continue, but this time wants a slightly different type of data.

When work resumes, the supplier finds that the new instructions make tasks more time-consuming. The cost estimates previously given to the lab no longer hold. So the supplier either cuts pay or finds ways to make workers move faster.

A new batch of data is delivered, work pauses again. Perhaps the lab adjusts its data requirements once more, or decides it has enough and ends the project; or simply switches to another supplier. Maybe the lab now only wants organic chemistry experts, and everyone without that background is removed from the project. Further down the line, the need might shift to biology data, architectural sketches, or curriculum design from elementary through high school.

To stay competitive, data companies arrange everything so that people are always on standby, while retaining the freedom to drop them at any moment. "Every supplier will have some arrangement that essentially amounts to not making real commitments to people," said a veteran employee at a large data company.

These changes often come with little advance notice even to the companies themselves. Sometimes because the AI developers weren't sure what data they needed in the first place, sometimes because they're shopping around for the best deal.

"They want to keep us in the dark," this employee continued, "so we inevitably keep data contributors in the dark too. Then a procurement falls through, and a thousand people you've trained and built relationships with are saying: 'What the hell? Why is there no work?' From an operator's perspective, that feels terrible too — but obviously, it's far worse for them."

At the bottom of this supply chain, workers exist in an environment of extreme instability and intense competition. Especially strict non-disclosure agreements prevent them from accumulating any credentials or relationships that could carry beyond a given project.

"All the power is on one side, because they can't talk about their work," said Matthew McMullen, a strategy and operations executive who has worked in the industry since the mid-2010s autonomous vehicle boom.

"You can't leverage your experience in the market, and the labs benefit from that. That silence is like their pricing power. It's that silence that lets them extract massive amounts of information from people without giving them the power to object, organize, or start their own companies. As long as these raters can't prove what they've done, they can't demand compensation matching their value.

"The only way people can make demands is by proving they can keep stepping up and taking more work. The only power they have is to keep working, to get back in line."

And that's exactly what they do. When a Mercor project ends, managers typically post links to other projects on the platform, encouraging people to apply.

"But it's thousands of people applying at once, so it's like throwing your application into a hole and hoping to hear back at some unknown moment," Katya said. While waiting, workers go register on Handshake, Micro1, Alignerr, or any of the other data supplier platforms constantly emerging.

These companies are always hiring. Like Mercor, many use AI interviewers and automated assessments, so there's no incentive to limit interview volume. Mercor offers hundreds of dollars in referral bonuses, leading some people to promote the company so aggressively that multiple Reddit communities have banned mentions of it.

Katya applied to dozens of positions and got three. That ratio is not unusual.

Companies bear no cost for over-hiring. Since workers are nominally independent contractors, companies don't have to provide paid leave, rest breaks, health coverage, overtime pay, or unemployment benefits.

Keeping these people waiting costs the company nothing. The more people who have been vetted and are waiting to start work, the more it ensures they will rush to complete tasks before others. All these factors combine into an arrangement where employers can turn the labor tap on or off like a faucet.

(When responding to interview requests, Mercor spokesperson Heidi Hagberg stated: "The nature of this work is project-based contract work, which means it can be delayed, paused, or ended at any time, especially when client project scope and requirements change." She also said that many workers' complaints "center on a misalignment between expectations of project-based work and full-time work.")

All of this will eventually pass.

If you're fast enough, lucky enough, happen to have the right specialized skill set, and can keep hitting each platform's unique and opaque productivity metrics, you can make decent money. I interviewed a playwright earning $10,000 a month, and a polymathic chemistry expert who, at various points, had taken gigs demonstrating poker strategy and singing for AI.

But even then, there's no escaping the feeling that it will all eventually disappear.

Because producing training data means contributing to your own obsolescence. Even as the number of people doing data work continues to grow, each individual gig lasts only until the machine successfully mimics it. Humans spend years mastering expertise; sooner or later, they'll have no skills left to sell.

One worker with a master's in linguistics had spent a full year on steady rubric-writing work. But by late 2025, he found it increasingly difficult to stump the models. No matter how obscure the theory or which Indigenous language he threw at them, they found the right papers. Where he used to turn in three or four rubrics a week, he was now lucky to complete one.

Everyone else on the project was experiencing the same thing, so when it ended, he wasn't surprised. Their expertise had been extracted.

In the past, he'd always found a new gig. But now, looking around, the only openings he saw were for medical specialists, HR managers, and teachers. He'd been without work for five months and didn't know what came next.

"Worse than driving for Uber"

These platforms recall Uber and Lyft a decade ago. But in some ways, these workers are worse off: even with advanced degrees, they're more easily replaced. When policymakers discuss how to respond to AI automation, the focus is mostly on what to do when AI eliminates entire professions en masse. Perhaps that will happen. But another possibility is that specific tasks get automated while humans are reassigned to other parts of the production pipeline — some fixing mediocre AI outputs, others writing rubrics to improve them.

Much of this work is inherently intermittent, which means it will be done by independent contractors whom existing regulations barely protect.

MIT economics professor Daron Acemoglu, who studies automation, compares their situation to that of weavers. Before the Industrial Revolution, weavers "were like the aristocracy of labor." They were self-employed craftspeople who controlled their own time.

Then came the power loom. To survive, they were forced into factories where, under close managerial supervision, they worked longer for less.

The problem wasn't just that technology took their jobs — it was that technology enabled a new form of labor organization that concentrated all power in the hands of capital owners. Capital owners made work a nightmare until workers organized and regulations set boundaries.

Early labor conflicts are already emerging, mainly in California, which has among the strictest rules for determining platform workers' employment status. In the past six months, Mercor has faced three class-action lawsuits. (Surge AI and Scale AI have faced similar suits; Scale AI is settling.)

All allege that the companies exercise "extreme control" over workers while misclassifying them as independent contractors. This is "a completely new kind of work" that requires company training and can only be performed on company platforms.

Workers know remarkably little about what they're actually doing. One lawsuit filed last December alleges that a worker, after accepting a Mercor project, was required to record himself reading aloud from scripts containing explicit sexual descriptions. Once he discovered the task's nature, quitting mid-project risked account deactivation, forcing him to "choose between getting paid and being humiliated."

Glenn Danas, a partner at Clarkson Law Firm, says these companies recall Uber and Lyft a decade ago. His firm is suing Mercor and several other data platforms. But in some ways, these workers are worse off: even with advanced degrees, they're more easily replaced.

Uber drivers must be physically present in a city to work there, so they can organize locally and push for regulation. If data workers try the same, companies can simply hire elsewhere, where people will accept lower pay.

When Mercor cut the Meta project's pay to $16 an hour — below minimum wage in California and several other states — workers in those places kept working because they needed the money.

At least one supervisor acknowledged this, writing on Slack: "While we won't actively recruit in states where the minimum wage is higher than the project rate, if you're already on this project and willing to accept $16 an hour, we want to enable you to continue."

Acemoglu says companies can pit workers against each other, forcing everyone to rush to sell their data before someone else undercuts them. If this dynamic takes hold, entire occupational groups could find themselves in a race to the bottom.

"We may need something like unions, some form of collective ownership, to prevent large companies from using simple divide-and-conquer strategies to suppress data prices," he says. "Without a legal foundation suited to this data economy, many of the people producing data will be underpaid — to use a weightier word, exploited."

Katya was among the thousands invited to join the Nova project at $16 an hour. The rate shocked her. "I felt like it was an experiment by Mercor to see how low they could push pay without compromising the data they were getting," she says.

The main project she'd been working on had been paused for weeks — it might resume tomorrow, or never again.

Eventually, she decided it wasn't worth it. She applied for a job at a local coffee shop. This wasn't the career transition she'd imagined in grad school; she just hoped being a barista would be more stable.

"At least at a coffee shop, making minimum wage, you have friends to talk to, a boss who at least pretends to care. You have some security, knowing what your weekly hours will look like," she says.

But then her phone dinged. One of her projects was back online.

By Josh Dzieza

Source: You Could Be Next[1]

[1] You Could Be Next: https://www.theverge.com/cs/features/877388/white-collar-workers-training-ai-mercor

Recommended Reading

Join Our Membership Group