When a startup closes, who gets its old data?
Brokers are selling code and work records from failed startups to AI buyers, MIT Technology Review reports. The records may include chats and other material people shared for a different purpose. A sale may pay creditors, but it does not answer every privacy or consent question. Broker deal counts are reported claims, not an audit of the whole market. No public test shows how much these records improve an AI model.
What the source reports
MIT Technology Review reports on brokers who offer old startup data to AI training buyers. A failed firm may still hold code and work records. Some records can include chats and other details that people shared while doing their jobs. A sale may raise money for creditors. Yet the people in those records may not have expected their words to train an AI model. The report raises questions about privacy, consent, and what a buyer will disclose. Broker examples show that a market exists, but they do not count every deal. No public test in the report shows that the data makes a model perform better. Nor does the report say every sale is against the law.
Original source
- Title
- AI labs are buying data from dead startups to train their models
- Author
- Matthew King
- Publication
- MIT Technology Review
- Date
- Tuesday, October 13, 2026