OpenAI Will Start Training on Data It Used to Keep Off-Limits

Starting November 29, OpenAI will begin training its models on user sessions from Daybreak Blue, its Trusted Access program for cybersecurity. These sessions were never used for training before, even if you had the "improve the model for everyone" option turned off. Users pushed back, and in response the company said it would add an explicit opt-out.

This reminds me of another case. Two mathematicians, Tristan Buckmaster and Levent Alpöge, posted draft work on solving the Navier-Stokes equations in Codex, OpenAI's code generator. Afterward, OpenAI published a solution to the same problem and said it had not used the mathematicians' work. An explicit opt-out, added only after complaints, is not a real opt-out. It is just a checkbox you tick once, and the company can change what that checkbox actually means at any time. If you hand your sensitive data to another company's API, odds are it ends up in their training data eventually.