Lately a new kind of pitch has found its way into my inbox. “Not sure if you’ve heard,” writes someone named Nick, “but AI labs such as Anthropic and OpenAI are paying media companies such as Wired $350K to $1.8M as a one time fee for a copy of their old work records (documents, emails, slack messages, tickets, etc.) for training better models.”
Another email with a similar pitch (adding call transcripts to the list of desired information) raised the stakes to a possible $2 million. “Public data is exhausted, and the labs are spending over $45B on private data to build better models,” writes someone named Cora.
While I was aware that AI companies are buying expert data from various sources, there hasn’t been that much attention paid to it. But once I looked into it, I discovered that this business is way bigger than I suspected—and according to its practitioners, it could have as much impact on workers and society as the data centers that are dominating the news these days.
The companies that contacted me are part of an industry boom in data acquisition that’s been quietly buying the workflows of sizable companies and is still so hungry that it’s bleeding into our inboxes. A freelance writer recently wrote about her experience with a company called Handshake, which offered her $1,000 for her work on a single story. (She didn’t sell.) A representative from Handshake confirmed the offer to me: “We license professional documents across a lot of fields so the labs can test and train how models handle real knowledge work,” he said. “Journalism is one of those fields.” Handshake tells me that a big part of its business is buying graduate students’ papers and coursework.
An entrepreneur friend of mine tells me he’s been getting mail from multiple companies whose business is “helping startups wind down their companies,” as one email from a company named Sunset explained it. “We handle the full process, scrubbing and anonymizing the data so the company people and any identifying details are stripped out, then brokering the deal with the labs.”
Those newer data acquisition startups are hoping to achieve the quick success of the dominant companies in the field, which themselves are only a few years old. Industry leaders Handshake, Micro1, and Mercor are snapping up databases as if it were a modern gold rush. The industry leader seems to be Mercor, founded in 2023; last year its three cofounders became billionaires at 22. Like all the companies, it claims to have partnerships with just about all the major AI labs. (The response from the companies themselves was more circumspect: Anthropic acknowledges that there’s a big market for data, but told me that it doesn’t buy workflow from live companies like WIRED. OpenAI didn’t come up with a response before publication.)
In an interview, Mercor CEO Brendan Foody told me that the market is poised to explode as the AI industry seeks dramatically more data. Mercor originally started as an AI recruiting company, but as the team got to know the players, it learned that the big labs were frustrated by the quality of the data. Generally, the industry had been hiring experts to simulate actual workflow, but that didn’t capture the messiness of an actual operating business. He urged me to consider that Big AI has two bottomless needs: compute and data. While the industry is very publicly making giant investments in the former, data gathering is just as important.
Foody is ready when I bring up privacy concerns. When a business streams the data to Mercor, the company strips out any personal identifying information, or PII. This de-identification can be verified, he says. Like all the people in data acquisition I spoke to, he felt that technology has solved the privacy issues of buying all that information.
Not surprisingly, Micro1 CEO Ali Ansari agrees that the data-buying market will soon approach that of AI’s compute infrastructure. He says that today, labs, enterprises, and robotics companies pay about $40 to $50 billion a year to get data. (I was unable to verify this.) “In the next two to four years, this is going to be over a trillion dollars a year,” says Ansari, who founded the company in 2022. Micro1 uses the information it buys largely to improve AI models it builds itself, but it also sells to big AI firms.
“Pretty much everybody’s trying to get into the data space now because there’s so much demand,” he says. “If you look at the YC startups, about half of them are data companies.” (Actually, it’s more—in the most recent YC batch, 40 of the 65 launched companies are classified as data startups.)
The Wild West aspect to this quest for fresh data got rare exposure when a bankrupt Spirit Airlines put its digital assets up for sale. The corpus included 80,000 email accounts, 500 million internal messages, 20.6 million Microsoft Sharepoint files, and a million time-card records. Google seemingly got the prize with a $10 million bid, but then Micro1 topped it by $2.5 million, setting off an argument over who the winner was. Presently the sale is on hold, for a reason that calls into question the entire practice of selling off human workflow: The flight attendants’ union has charged that the sale involved their private information.
The union’s objections cast a light on the dark side of this gold rush. Despite the claims of virtually all the data companies that no PII goes into AI models, critics feel that there is still danger of personal data leaking out. “It’s not clear that anonymizing can fully get rid of privacy issues,” says Mayu Tobin-Miyaji, general counsel of the Electronic Privacy Information Center, which has filed an amicus brief on behalf of the flight attendants. “Emails and Teams messages contain unstructured data, which is harder to root out. There may be messages where employees talk about doctor’s appointments or family issues.”
“We haven’t set the norms and the standards for data acquisition, and it’s not being done appropriately ” says Gordon Ritter, a venture capitalist who has taken on the issue as a cause. “The AI industry is moving far faster than our ability to actually set the standards, and it’s partially intentional.”
Probably the most troubling aspect of the data-gathering movement is that, in essence, it enlists workers, not always with explicit consent, to contribute to an effort where the ultimate goal is to replace or automate the human work that goes into all those documents, messages, and emails. When Mercor says that its motto is “Organizing human intelligence to power the AI economy,” the unspoken message is that the economy of flesh and blood workers will have less work to do. As the president of the flight attendant’s union told an interviewer: “This is really only data that became valuable in the age of AI in order to train models to replace jobs and commodify other portions of working in a work environment.”
When I ask Micro1’s Ansari whether one day the default for corporations will be selling their workflow data to big AI companies on an ongoing basis, he answers affirmatively. “I think it’s going to become more common than ads as a revenue line item for companies,” he says. I note that there is already considerable pushback to AI, and that people might not be happy with the idea that everything they do at work gets funneled into the maw of a big AI model.
“I think it’s a fair point,” says Ansari. But, in his view, such objections will be manageable. He suggests that workers will understand that the fruits of their labor will result in better AI models, which will make their lives easier. “You have to think of the upside, the abundance that AI can result in.” (I’ve found that the future abundance that AI will supposedly produce is a handy justification for all sorts of compromises.)
I did follow up on one of the cold pitches I got, from a woman whose email domain was gorexen.com. A “gorexen,” by the way, is a fusion of two Pokemon critters that “siphons the body fluids of prey through its thin, tubular mouth.” She set up a meeting with me with a two-year-old company called Tenkara (the name seems to be associated with a type of Japanese fly fishing). I found myself talking to “head of commercialization” Jonah Stillman, a young startup-looking dude sitting in an office in San Francisco.
I began peppering him with questions. Tenkara’s quest for data seems to be to boost the quality of its AI model to run office operations for manufacturing companies. (How WIRED’s archives would help is a mystery.) He confirmed that a big part of his job is doing calls like this to snap up data. I told him that the set of our meeting gave me a Glengarry Glen Ross vibe, but he wasn’t familiar with the reference.
I am, of course, not authorized to sell the documentation of WIRED’s editorial process to an AI company, and I did not risk my job by offering Jonah and his company the goods. I also wouldn’t be too pleased if my bosses made the sale.