AI labs are buying data from dead startups to train their models

MIT Technology Review reports on a growing market for the digital remains of shut-down companies. Wind-down firms like SimpleClosure and Sunset now broker sales of code and workspace data, including emails, chats and documents, to foundation AI labs, training-environment vendors and big tech companies building their own models.

This matters because the data isn't really the company's in any human sense. It's years of employees talking to each other. The piece describes it as increasingly routine, and there's no federal law banning these sales, according to an Emory law professor quoted in the story.

We understand why this market exists. Labs have hit the limits of public and synthetic data, and real workplace archives are messy in the ways agents need to learn from: disagreements, judgment calls, real dates. Founders whose companies failed get a little money back. In the piece, one founder says he netted roughly $25,000 for his code base. That's a real benefit, and we don't think the answer is to pretend the data has no value.

But the consent problem here is serious. Workers accepted monitoring because it helped them do their jobs. Ifeoma Ajunwa's point lands: once the company is gone, that justification is gone too, yet the archive gets sold anyway, and the proceeds typically flow to senior creditors and investors, not employees. That's a quiet transfer of value from workers to whoever holds the cap table.

The privacy side worries me more than the vendors seem to admit. Sellers and buyers say they scrub personal information and exclude HR records. Attorney Greg Ewing's examples are the right ones to sit with: doctor appointments, school pickups, date-night plans tucked into calendars and chats. Automated scrubbing catches the obvious fields. It's much worse at the human details buried in ordinary conversation.

And nobody has shown yet that this data meaningfully improves models. The piece says no company has released tests comparing real-world data training with the alternatives, and buyers quote wildly different prices based on research theses. So we have a fast-growing market built on an unproven premise, with weak consent and uncertain privacy protection. That's a familiar pattern, and it rarely ends with the people in the data coming out ahead.

Our read: this should come with express employee consent and a clear record of who bought what. If the industry won't do that on its own, regulators and courts will likely do it for them, starting with the next big bankruptcy sale.

SimpleClosure added data sales in April and says its Asset Hub has handled over 150 deals, with 47% of its deals now including code or data. In August, as part of Spirit Airlines' bankruptcy, Google outbid Mercor to buy the carrier's data for $10 million, including 100 million emails and 500 million Teams messages; Spirit's flight attendants' union objected on privacy grounds. Bloomberg reported in September that SpaceX's AI unit is weighing similar purchases.