The Journal. - How AI Is Being Trained to Do Your Job

Episode Date: June 4, 2026

There’s a new gig economy that involves training AI to do white collar jobs. And one company, Mercor, is leading the charge. The training startup hired 30,000 contractors just last year to help AI companies get their models trained up on sophisticated roles. Ryan Knutson spoke to one former Mercor contractor about the job and WSJ’s Katie Bindley lays out the tactics that have landed the startup in hot water. Further Listening: The ‘Class of AI’ Enters the Workforce AI Is Coming for Entry-Level Jobs Sign up for WSJ’s free What’s News newsletter. Learn more about your ad choices. Visit megaphone.fm/adchoices

There's a new type of gig work that's growing rapidly right now. It involves talking to AI and teaching it how to do your job. There is this new sort of category of jobs that basically involve people sitting in front of a computer and helping to train AI to perform some of the things that they are, you know, experts in or what they used to do in their careers pretty much. Our colleague Katie Bindley covers tech, and she's been reporting on a startup that's doing this kind of work. And it's hiring like crazy. It's called Mercor. The best ways to sort of get insights about this type of work is to just look at the job postings that are available. They put out a gazillion postings on LinkedIn.

Like, can we look at some of their LinkedIn posts? Ooh, let's. Okay. So you'll see tons of posts. Start your month with work from home. explore flexible remote opportunities with Mercor and work from anywhere, apply here. It'll basically say, like, we are looking to hire someone with experience as, you know, an investment banker, entry-level analyst or whatever, and it lays out specifically what they're looking for,

and I think the idea is to help the AI perform the duties of those roles. Are they ever looking for podcast hosts? I've seen ones for like voice actors and stuff. I mean, the range of job postings, it's like hard to think of jobs that, they're not hiring for? I mean, they were looking for dermatologists, radiologists, therapists, like accountants, consultants, actors, I mean, screenwriters, all kinds of things. Is there a metaphor that you could use for like this phase of AI training? Like, the first phase of training AI is like reading every book in the library. Yeah. And so now going to sit down with a

tutor maybe? Yeah, I think actually a tutor is like a really good way to put it. It's certainly like moving us toward the direction of AI being able to handle more and more complex sort of white color work-related tasks. Welcome to The Journal, our show about money, business, and power. I'm Ryan Knudsen. It's Thursday, June 4th. Coming up on the show, the new gig work that's training AI to take your job. Merkore was founded in 2023 by three college dropouts. It's become a $10 billion company in less than three years.

It's basically a middleman between large AI companies and humans who can help train large language models. AI companies need a lot of data in order to improve their models. And the gist is that most, if not all, of the publicly available data has already been used. So in order for the AI to get smarter, if you will, it needs fresh sources, fresh materials to train on, and humans turn out to be very good sources of that. Mercor has clients. Their clients have included some of the biggest AI companies, including Open AI and Anthropic. And so they're serving as sort of this go-between, and Mercor hires the contractors. And so then the contractors are the ones who are helping to improve the models.

One person who's worked as a contractor for Mercor is Carolina. I'm Carolina Perez-Sands. I'm originally from Spain, from Madrid. I grew up in three languages, right? So my parents are from Spain, but we lived in Lisbon. So my friends were Portuguese. And then I went to a French school. So I grew up in three languages simultaneously.

And I was always very interested in words. Carolina came to the U.S. over a decade ago, went to graduate school, and spent a big part of her career working as a speech and language pathologist. So what I did was I helped people for whom voice was their main job instrument, if you want. So I helped journalists, reporters, TV and radio presenters, actors, actors, singers, singers, lawyers, everybody. who needed to have a voice that sounded powerful. Recently, while trying to grow her business, she started promoting her language and writing skills on LinkedIn. And it led her to a job she wasn't expecting, training AI.

This was March 2025. I was to attract clients on LinkedIn, posting, etc. And because I had on my headline that I spoke Portuguese, Portuguese, Portuguese, Spanish and English, somebody from an agency called Merckor approached me because they were looking for Portuguese-speaking writers. The pitch is you speak Portuguese? Do you want to apply for writing, analyst, bilingual in AI? And I said, oh, yeah, having a title that says,

writing analysts, it sounds fun. What attracted me was to have the writing analyst in my top title. In reality, I always wanted to be a writer, but I thought I could never make a living being a writer, so I never tried.

Mercor told Carolina that she would be assigned projects with major AI companies, and that she'd be using her skills to train their large language models. Like most Mercor contractors, Carolina signed an NDA that prevents her from revealing the name of the client she worked for.

I thought, well, okay. Okay, why not? I'm going to be paid. The pay was not bad at all. How much was it? $45 an hour. How did you feel about that as a job to just train AI systems? In the beginning, I felt great because I thought to myself,

oh, I'm in the cutting edge of technology. I know things that most people don't and things like that. And it was really interesting. For contractors, being on the cutting edge of technology means trying to make the AI models good at the things they're good at. For Carolina, the task was to help it improve language and writing skills. Her first project was simple. The client asked her to evaluate how well the AI was responding in European Portuguese,

as opposed to Brazilian Portuguese. The task in the beginning was to make sure that the written responses that the models gave included phrases that were commonly said. And because I was doing European Portuguese, if the response I got had a lot of Brazilian Portuguese type of phrasing, I had to correct it. I would click on a task and I would see the prompt and then the response. And I had to say, well, this response doesn't match the prompt well

because they were speaking with a lot of Brazilianisms. So I had to adjust and say, no, in Portugal, we don't speak like this. We say it like that or whatever. Beyond just catching simple language errors, Carolina also worked on projects that involved teaching AI to write and speak more like a human. We needed to create prompts to give us creative writing out. So if I want to write a memoir, I remember one of the prompts that I wrote was write the opening scene for a memoir where a woman lives in Miami Beach. And it makes so obvious choices that it said,

yeah, I was lying on the floor and there was sand everywhere. Like the fact that you live in Miami Beach doesn't mean that you live on the beach. And there's going to be sand in your apartment. Exactly. So those were the things that I needed to flag. And sometimes at the first reading, the first pass was, yeah, okay, it's not bad. And then, like, wait, this doesn't make any sense. Why is she lying on the floor?

Yeah. As the projects went on, and she kept training the AI to sound better, Karolina said that she and her fellow contractors got really good at catching all the cliches and patterns that the AI needed to work on. We had a Slack channel. There were a lot of screenwriters, playwriters, literature professors, historians. And we started to notice AIisms. Like, oh, AI does this all the time.

What kinds of things is an AIism? At some point, everything happened on a Tuesday. Oh, really? Because they want to, they have learned in creative writing, 101 for AI models, they have learned that you need to write specific things and land everything in real life. So they say, it was a Tuesday and everything was a Tuesday. The most average of days. Exactly. And that's exactly what I thought. Like, why would Tuesday be more average than Wednesday? Like, it was interesting. Our colleague Katie spoke to a lot of contrasts.

tractors like Carolina. And she says they have mixed feelings about this type of work. Not everyone is like so psyched to be doing this. I think they, in some cases, they just can't find work. And this is a way to make some money. And there also is, there was this sort of sense of inevitability that I think they all felt of like, I mean, look, the whole question of like how many jobs will AI automate away super nuanced, complicated topic that nobody really knows the answer to. Right. But among these workers, they kind of feel like, okay, writing's on the wall, that at least some of the stuff I used to do is going to be automated. So like, I'm not going to be able to prevent that. So why don't I just like make some money kind of on the way out, I guess?

It's an interesting thing to do, I feel like, as a gig, if you're out of work in that industry, because it's almost like you're burning down the building behind you, you know? Like, if I can't get a job in this industry, nobody can. Yeah, I mean, people I spoke with. that is much. Murkhor said in a statement that training AI is becoming one of the fastest growing categories of work in the economy. The company said, quote, we are paying experts more than $3 million a day for that work. The company added that every major technology shift has looked like a jobs crisis, but the consistent pattern is that it creates new categories of work. Coming up,

how some of Mercor's tactics have landed it in hot water. One thing that happens at Mercor that can make contractors uncomfortable, is that they're sometimes asked to share prior work. The trouble is, much of their prior work could technically be owned by a previous employer. Their responsibility to the client is to be helping to make these models capable of performing more sophisticated work, and that in order to do that, in that sense, what you need is sort of people with prior work materials. The issue there is that there is a ton of prior work material people have, but they don't tend to own the rights to it. And so then you get into some of the stuff that I reported on where Mercor was approaching, in this case it was visual effects artists. They were messaging them on

LinkedIn and saying like, hey, we're looking to purchase high-end production work such as, quote, 4D physics scenes with camera data, depth and motion point tracking. And the guys I spoke with who had been approached, were very taken aback by this request. Because even though they're contractors, the work they had done was for major studios. They said they sign NDAs that are like the thickness of a book. And everything's under NDA. It's all protected IP. So they were very taken aback by Mercor offering to purchase this data and work product

because they were like, dude, we definitely don't own this stuff. What does Mercore said about this? So Mercor said that they do not buy intellectuals, property and that they license content directly from individuals and only when they own it. They also said that they do not ask for, nor do they want material owned by a current or past employer. So they're saying we only want stuff that people actually have the rights to license us. Another issue came to light after Merkhor was hit by a big data breach. They suffered a large data breach recently, which, to be fair, many, many, many,

startups suffered a breach around the same time. After the breach, Merckor was hit by at least seven class action lawsuits. They alleged that Mercor exposed contractor information, ranging from recorded job interviews to facial biometric data and screenshots of workers' computers. Within the claims of the suits were a lot of really interesting allegations about the ways in which the company allegedly acquires the data that it uses to serve its customers. So for one thing, this is sort of interesting, is that the contractors, their computers, while they're doing their work, Mercor has software that takes screenshots of their work.

And so among the allegations were that those screenshots were shared with the clients and sort of part of the data that they were handing over to the clients. So essentially like watching the employees work and seeing what they're doing on the computers is a form of data that's being used to train AI potentially. That's what the allegations. are in the suit. And, you know, Mercor makes the point that, like, we tell people that you are only to be working on your Mercor-related projects when you're billing for time. So, like, if you have two monitors, you should only have your Mercor-related work up on either screen at any time.

Mercor said they strongly dispute the speculative claims in these lawsuits, and they look forward to presenting the facts at the appropriate time and place. And they also added that they take the privacy of their customers, contractors, employees, and those they interview seriously and comply with relevant laws and regulations. Yeah, so that was their response to some of the reporting and the allegations included in these claims. Still, Mercor is continuing to attract contractors who want to get paid to train AI. Last year, the company hired 30,000 of them, and it's also attracting a lot of investors, too. The company, they raised quite a bit of money, and it was in the fall that they had a new funding deal

that valued them at $10 billion, and that actually was five times what they'd been valued the February prior to that race. So they've been one of those sort of, you know, quickly rising companies. The startup also has competitors now like Handshake AI, Micro One, and Surge,

that connect AI companies with contractors looking for AI training gigs. How long do you think companies like this will need to exist, though? I mean, isn't the whole point to train AI? so that it's capable of doing everything? So what won't theoretically, Mercor succeed so much

that it no longer has a reason to exist? I mean, I think these are big, important questions. There's so much uncertainty with respect to what AI's capabilities will be. What will the cost limitations be? The resource limitations. I mean, there's... I think sort of my takeaway is that the ambitions of the companies are limitless. This is basically what happened at a smaller scale to Carolina.

She only worked in Merk-Corps for about a year. After her first assignment, that one where she had to flag Brazilian versus European Portuguese, she was surprised at how fast the AI was learning. That's when I worked myself out of her job because the model learned so fast. So with all the data that we were inputting, like in a week's time, you didn't have to correct anything. Like the prompt and the response were perfect, nothing to correct. And so you look at it and be like, oh, yeah, that's pretty good.

That's pretty good, but on the other side, I didn't have a job anymore. After a few more projects with Mercor, she started to have more negative feelings about the work and about AI itself. I started to think my job is actually making this more of a monster. We're like just putting out pollution out there. And what is it doing to the culture? to the society. I thought it was really like a waste of talent, a waste of time, not for me, but a waste of

collective time. Like, we're not connecting people to people. So yeah, I started to feel very disillusioned. Tell me about when you made the decision to quit. Well, actually, because so this is the thing. I needed the money. And I decided to quit when the money was really, like, the last project I did, instead of $45 an hour, they would pay $35 an hour. And then the very last one that I did, they were paying by the task. And the pay was something around $20 per task.

And I thought, this is ridiculous. So I did five tasks and I quit because I had. I didn't think it was worth it. Recently, Carolina got a new job. She now goes door to door talking to humans as a salesperson. It's still an open question as to whether AI will be capable of doing people's jobs and to what degree. But what this story tells me is that these AI companies are coming for you. Like, they're trying.

Yeah, totally. You know, whether or not they're going to succeed or not is an open question. but like they're trying to make AI good enough to do it. Yeah, no, I absolutely. That, you know, that's definitely was one of my takeaways as I've been reporting on this and talking to all these people about the type of work they do. I don't really see evidence that they're like,

well, no, we shouldn't automate that type of white collar work. It feels pretty sweeping. That's all for today. Thursday, June 4th. The journal is a co-production of Spotify and the Wall Street Journal. Additional reporting in this episode by Angel Al Young. Thanks for listening. See you tomorrow.