The Jobs AI Is Creating Are Worse Than the Ones It's Killing

The story behind the data

Two days ago, I wrote about Anthropic's research showing 75% of programming tasks are already being done by AI. Financial analysts, technical writers, and market researchers all landed in the top 10 most exposed occupations.

That piece was about the numbers. This one is about what happens next.

The Verge and New York Magazine just published a deep investigation into the growing industry of displaced white-collar professionals being hired to train the AI that replaced them. Lawyers, screenwriters, graphic designers, and scientists, all funneled into precarious gig work producing training data for the same models that took their jobs.

The company at the center of the story is Mercor, a data-labeling platform valued at $10 billion. It was founded in 2023 by three 19-year-olds. Last year, it made them the world's youngest self-made billionaires.

Their workforce: 30,000 professionals per week, writing rubrics, grading chatbot outputs, and producing "golden answers" that teach AI to mimic their expertise. Scale AI, a competitor, claims 700,000 workers.

One Emmy-winning documentarian described it this way: "I'm being handed a shovel and told to dig my own grave."

The mechanics of the trap

Here's how it works.

An AI lab discovers its model is weak in a particular domain: chemistry, legal analysis, financial forecasting. It pays a data vendor like Mercor to find professionals in that field. Those professionals write examples of ideal work, create detailed scoring criteria, and try to stump the model with tasks it can't do yet.

Then they wait. The data goes back to the lab. The project pauses. Maybe it resumes with different requirements, shorter timelines, and lower pay. Maybe it doesn't resume at all.

A worker with a master's in linguistics found steady work for a year writing rubrics. Then in late 2025, he noticed the models were getting harder to stump. Any obscure theory or Indigenous language he threw at them, they'd find the right papers. Instead of three or four rubrics per week, he was lucky to get one. Then the project ended. His expertise had been extracted. He hasn't found work in five months.

This is the defining feature of AI data work: every task you complete makes you less needed.

What "reskilling into AI" actually looks like

The narrative has been that displaced professionals should "pivot to AI." The Verge piece shows what that pivot looks like in practice.

• Pay that drops every cycle. One project started at $21/hr. Mercor ended it, relaunched the exact same job under a new name, and cut the rate to $16/hr. That's a 24% pay cut overnight. Workers reported that demands increased, timelines shrank, and pay decreased as every project continued.

• Surveillance tracked to the second. Software called Insightful monitors everything workers do on their computers. Minutes without typing trigger a check-in. "Unproductive" time gets deducted from pay. One worker started turning off the monitoring software and working for free when he went over the target time, because falling behind meant getting fired.

• Work that arrives at 7PM on a Sunday. Tasks drop in Slack without warning. Workers describe "jumping on them like piranhas" before they run out. A screenwriter with streaming credits said she spends no time with her kid and screams at him when tasks come in during dinner. "This work is turning me into a fucking demon," she told The Verge.

• No notice, no explanation, no recourse. Projects end without warning. On one platform, workers call it "the dash of death." You log in and your dashboard is empty. No email. No explanation. You just don't have a job anymore.

Three class-action lawsuits have been filed against Mercor in the past six months, all accusing the company of misclassifying workers as independent contractors while exerting "extraordinary control" over their work.

The desperation math

This doesn't happen in a vacuum. It happens because people are desperate.

The U.S. economy lost 92,000 jobs in February. Long-term unemployment (27+ weeks) is at 1.9 million, up 27% from a year ago. Hiring is at its lowest point in decades outside the 2008 crash and the pandemic.

Handshake, the early-career job platform, found that postings declined 16% year-over-year while applications increased 26%. Their response was to launch a program connecting job seekers with AI training data roles. "As AI reshapes the future of work," they wrote, "we have the responsibility to rethink, educate, and prepare our network to navigate careers and participate in the AI economy."

That's the framing: this is the AI economy. Participate.

Meanwhile, people are paying thousands of dollars to get recruited. $1,500/month plus 10% of first-year salary to services that mass-apply on their behalf with AI-generated messages. $400 on Fiverr for 50 applications that lead nowhere.

When you're six months into a search with no income, a $45/hr Slack ping at 7PM sounds like a lifeline. The article makes clear why people take these jobs. It also makes clear why they shouldn't plan around them.

What this means for your search

This is not a story about a bad company. Mercor, Scale AI, and Surge AI are all doing versions of the same thing because the economics demand it. The AI development cycle creates inherently intermittent work. Projects pause when data goes back to the lab. Requirements change. Budgets shift. Workers are disposable by design.

The takeaway for your search is more specific than "avoid data labeling."

• Distinguish real AI roles from extraction jobs. Companies that are building AI capabilities need people whose judgment compounds over time. People who can architect systems, interpret outputs, and make decisions the model can't. Companies that outsource to data platforms need your expertise exactly once, packaged into training data, then they're done with you. The difference matters. Ask in interviews: "Is this team building AI tooling, or are you a customer of someone else's?"

• Follow the capital, not the gig listings. Last week, Cerebras raised $1B, Waabi raised $1B, ElevenLabs raised $500M. These are companies with fresh funding and aggressive hiring plans. They need domain experts to build products, not to label data. The roles they're creating look nothing like a Mercor project.

• Your expertise is worth more than $16/hr. MIT economist Daron Acemoglu compares the current moment to weavers before the industrial revolution. They were the "labor aristocracy," independent artisans in control of their own time. Then machines came and they went from artisans to factory workers with no power. The difference this time: you can see it coming. Don't sell your domain knowledge by the hour to a platform that will package it and discard you. Invest your search time in roles where that knowledge compounds.

• Watch for the "AI reskilling" framing. When a job platform tells you the future is "participating in the AI economy" by producing training data, they're offering you a position on a digital assembly line. The work is real. The expertise required is real. But it's structured so that you have no leverage, no stability, and an expiration date.

Why we built the opposite of this

The model described in this article works because workers have no information. They don't know who the client is. They don't know if the project will exist tomorrow. They can't see the instability coming. They sit in Slack waiting for a ping, and when it comes, they race to grab whatever's there before it's gone.

That's an extreme version of what the normal job market does to people. You apply to roles you can't evaluate. You interview at companies you haven't researched. You accept offers without knowing the team just lost half its headcount. The information asymmetry is the whole problem.

We built Kinship to fix that.

• Company signals before you apply. Every week, we scan 1,500+ companies for layoffs, AI-driven restructuring, funding rounds, and hiring freezes. When Block cuts 40% of its workforce or HP announces 4,000-6,000 "AI-driven" eliminations, that shows up in your job scores before you spend a minute on the application. One-click Deep Dive research pulls stability, funding, growth, culture, and leadership signals for any company you're considering.

• Fit scores that tell you why. Every role gets a 0-100 score with a clear explanation. Not keyword matching. It's based on what energizes you, what drains you, and what you actually need from a role. So you can see the difference between a company that needs your judgment long-term and one that needs your expertise packaged into a training dataset.

• Matches that come to you, on your terms. Daily email alerts with your top-scoring roles. You don't sit in a Slack channel at 7PM hoping for a ding. You open your inbox in the morning and see what's worth your time. The search works around your life, not the other way around.

• An AI coach that builds your leverage. Your Advantage takes your background, your strengths, and the specific role you're looking at, then builds a personalized competitive advantage playbook. Proof points, positioning, interview prep. The goal is the opposite of what data platforms do. Instead of reducing your expertise to a rubric, we help you articulate why it's irreplaceable.

The most valuable thing you have in a job search is your time. The people in this article are spending theirs producing training data at $16/hr with no guarantee of next week. That's not a career strategy. It's a conversion rate of zero.

The Anthropic data showed 75% of programming tasks are already exposed. The February jobs report showed 1.9 million people searching for 6+ months. This Verge investigation shows what's waiting on the other side for people who don't have good options. We're trying to make sure you do.