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

As AI development hits the limits of public and synthetic datasets, major labs are increasingly purchasing data from defunct startups to train their models. Companies like SimpleClosure and Sunset are brokering deals that allow failing firms to monetize their digital archives, including code bases, emails, and workspace communications. While demand for this 'operational exhaust' is surging, the practice raises significant legal and ethical concerns regarding privacy. Critics argue that even after scrubbing personally identifiable information, sensitive private details often remain in archived communications. High-profile acquisitions, such as Google’s purchase of Spirit Airlines' data during bankruptcy, highlight the growing scale of this market. Despite the potential for improved AI performance, the legal risks remain high, with some experts labeling such data as 'radioactive.' As the ecosystem for 'tech undertakers' grows, the industry faces a lack of transparency regarding the effectiveness of real-world data compared to synthetic alternatives.