AI labs pay data brokers to buy companies' emails and work files

A cold pitch reached Wired's Steven Levy. "AI labs such as Anthropic and OpenAI are paying media companies such as Wired $350K to $1.8M," one email read. The pitch asked for a copy of old emails, documents, tickets and slack messages. Another email raised the number to $2 million if call transcripts were added.

Wired reports that Handshake, Micro1 and Mercor are the named industry leaders. Each buys corporate workflow data and resells it to AI labs. Mercor was founded in 2023. Wired says its three cofounders became billionaires at 22 last year.

What is being sold

A representative from Handshake confirmed one offer to Wired. The company paid a freelance writer $1,000 for the work on a single story. The representative said Handshake licenses "professional documents across a lot of fields". The labs then use that data to train and test models. Graduate students' papers and coursework are another product.

Mercor CEO Brendan Foody told Wired the market is poised to explode. He said the big labs had been frustrated by expert simulation. Recorded workflow from a live business holds the mess a simulation does not.

Micro1 CEO Ali Ansari gave Wired a number. He said labs, enterprises and robotics companies pay about $40 to $50 billion a year today. Wired could not verify that. Ansari added the figure will be "over a trillion dollars a year" in two to four years. He also said 40 of 65 companies in the most recent Y Combinator batch are classified as data startups.

The Spirit Airlines case

A bankrupt company gave the clearest view of the market. Wired reports that Spirit Airlines put its digital records up for sale. The list included 80,000 email accounts and 500 million internal messages. It also included 20.6 million Microsoft Sharepoint files and 1 million time-card records. Google bid about $10 million. Micro1 topped it by $2.5 million.

The sale is on hold. Wired says the flight attendants' union has argued the records hold their private information. The Electronic Privacy Information Center has filed an amicus brief for them. Mayu Tobin-Miyaji is the general counsel of EPIC. She told Wired that "it's not clear that anonymizing can fully get rid of privacy issues". An unstructured message can carry personal detail a scrub misses.

What the labs say

Wired asked the labs in the cold pitches to respond. Anthropic acknowledged a large market for data. It told Wired it does not buy workflow from live companies like Wired. OpenAI did not respond before publication.

Every broker Levy spoke to made the same claim on privacy. They said personal identifying information is stripped before the data reaches a lab. Gordon Ritter, a venture capitalist, disagrees. "We haven't set the norms and the standards for data acquisition," he told Wired, "and it's not being done appropriately."

The market changes one thing for a team that runs a product on a commercial model. A supplier's training set is turning into a line a company sells on its own. That decision belongs to a buyer who has not had to make it before. The people whose work ends up in the set are rarely the ones signing.

Source

• AI Companies Really, Really Want to Buy Your Work Files and Emails, by Steven Levy, Wired