Source code and internal work records left behind by shuttered startups are being traded as data for AI training. As the view spreads that public web material and synthetic data alone cannot fully replicate real workplace environments, a new market is forming around the digital assets of failed companies.

MIT Technology Review, in its newsletter The Download on Oct. 28, introduced a growing number of cases in which AI developers are securing code, emails, workplace chats and documents from defunct startups.

Keelan Miskell (킬런 미스켈), the founder of a cryptocurrency startup, sold source code for about $25,000 as he wound down the company. He is also pursuing a deal to sell records left in Google workplace tools, Notion, Figma and Slack for about $15,000. The payment for the work records has not yet been finalised.

SimpleClosure, a firm specialising in startup liquidations that brokered the transaction, has been running a service selling code and workplace data since April. The company said related deals have surpassed 150, and 47 percent of all liquidation transactions included a sale of code or data.

Buyers are seeking to use complex collaboration and decision-making processes that occurred inside real organisations to train AI agents. Public comparative verification of how much real workplace data improves AI performance over synthetic data remains limited.

Concerns are also being raised over privacy and consent. Internal emails and messengers can contain sensitive information about employees and customers, and the authority and scope to dispose of such records for AI training purposes when a company shuts down may be unclear. Key issues remain, including who owns the data and whether stakeholders' consent was properly secured during the closure process.