OpenAI will start training on a security tool's data it once excluded

Starting November 29, OpenAI will use data from Daybreak Blue sessions to train its AI models. Until now, that data was kept out of training, even for people who had turned off sharing. An email to members of OpenAI's Trusted Access program confirmed the change.

What changed

OpenAI has a setting called "improve the model for everyone." Turning it off used to keep Daybreak Blue data out of training. That protection ends on November 29. The setting will no longer cover this data.

Backlash came quickly. OpenAI then sent a second email promising a separate opt-out. The material gives no details about how it will work or when it will arrive. Interactions flagged during safety reviews can still be used for safety purposes.

Why researchers are uneasy

The change has revived memories of a dispute involving mathematicians Tristan Buckmaster and Levent Alpöge. They uploaded drafts about the Navier-Stokes equations to OpenAI's Codex. OpenAI soon presented its own proof but denied using their inputs for training. The suspicion alone showed researchers what could go wrong.

What is not known yet

• How the promised opt-out will work
• Whether it will be on or off by default
• When it will be available

For anyone using AI tools on sensitive work, the lesson is simple. A setting you turned off may not cover a rule added later.