Employee Data Is Being Sold To Train Ai Models

If you're tired of censorship and surveillance, subscribe to Reclaim The Net.

AI labs' hunger for data to train their models has created a market in workplace records, including Slack messages, emails, and documents, and the code that employees write, but those employees are rarely told, or paid.

Business Insider reported that the data is "priced by the terabyte, anonymized, and sold to AI giants desperate for an edge." The report reveals that companies are getting six-figure offers that they must decide on within 48 hours, and that the market is full of data brokers, like Polyshares and Micro1.

A whole codebase was recently shopped around for millions – by somebody who had stolen it. Tech giants, who are among the most avid buyers, are not above mining the data of their own employees. Meta tried to do that, but suspended the plan after employees revolted and the data leaked inside the company. Elon Musk, on the other hand, told SpaceX employees that their data would be used to train the company's AI, and that this was something positive, as the AI would "inherit your thoughts and ideas and beliefs."

Polyshares says on its website that it "licenses the operating record a business produces by running, Slack, email, tickets, CRM, files, and supplies it to AI labs." The company counts as "operational data" that it collects and sells: "Slack, email, Drive, spreadsheets, CRM history, support tickets, design files, call recordings, accounting and ERP exports." And although it doesn't reveal which AI lab gets which data, cofounder Ryan Lock revealed that the offers being made to companies whose data is sold are similar to those made in the early days of the Texas oil boom: "They'll say, 'I'm offering you double market. I need an answer by Friday.'"

Meanwhile, the report said that Micro1 is another broker that is paying companies $100,000 to $2 million for access to their data, and last week announced it would spend $1 billion on company data over the next 12 months. Its CEO Ali Ansari believes that in the future, as much as 10% of US GDP could be spent on training AI models.

As for the people who wrote the messages, emails, and code being sold, Business Insider reported that employees rarely profit from, or even know about, "their Slacks, PowerPoints, and Google Docs being sold to the highest bidder." Emory University law professor Ifeoma Ajunwa said that there is little legal recourse for them, given that there is no federal privacy law. She added that it was "an ethically gray area."

Earlier this year, Meta tried to launch a program that would have tracked the keystrokes and mouse movements of its employees to train its AI models. Employees revolted. And even though CTO Andrew Bosworth said that the data collected in this way would be "tightly controlled" – it was found to be viewable across the company, and included private conversations and performance records that were supposed to be anonymized. Meta suspended the project.

Meanwhile, Meta is not the only giant that sees the data produced by its employees as a valuable resource. In August, Elon Musk told his employees at SpaceX that the company's AI, Grok, would be trained on their data. "In a way, it will be trained on you. It will inherit your thoughts and ideas and beliefs, and I think that's a good thing," Musk said.

The report also reveals that the market is rife with stolen data, and that the sellers are getting brazen. Nick Heiner of Surge AI said that he was approached earlier this year by a group who offered to sell him a large company's codebase for millions. But as it turned out, the group had no right to sell it – they had simply stolen it and were now trying to fence it.

"It became apparent that this group did not have the rights to sell this data," Heiner said. "They had illicitly obtained a copy of the codebase and were just going around to all the data vendors saying, 'Can I get a few million bucks for this?'"

Former gig workers for another AI training startup, Mercor, allege in a lawsuit that the company encouraged them to upload work materials without their employers' consent. Mercor denies it.

Then there is the case of Google trying to buy the data of bankrupt Spirit Airlines. That deal has hit a snag, partly because the airline's flight attendants are worried about how their records will be used. Google told Business Insider that it will make sure personal data is either removed or de-identified by an independent third party.

As for the data that was to be sold, we reported in August that it was 100 million emails, half a billion Teams chat items, and 175,000 personnel records. The roughly 17,000 people who lost their jobs when the airline went down are also in the trove, with "their old inboxes, their pay records, 1,092,000 time cards."

We also reported that the sale agreement was designed to preserve "referential integrity" of the data, which means that "a single pseudonymous employee could still be traced from a mailbox to a payroll file to a crew pairing." And while the agreement requires the data to be scrubbed of personal information, it lets Google choose the firm that will do the scrubbing, and pay for it.

In April, we reported that OkCupid had handed nearly three million users' photos to an AI company, and that a settlement with the US Federal Trade Commission (FTC) "carries no financial penalty."

In 2024, we wrote about a lawsuit against Photobucket for allegedly selling users' photos and biometric data to AI trainers without consent.

The only party that is not at the table in this multi-billion dollar market is the people who wrote the messages, created the documents, and generated the data that is being sold. The "anonymization" of this data is a promise made by those who sell it, and verified by a third party that the buyer pays for. Without a federal privacy law, the only thing that is needed for the data to be sold is the employer's signature.