Starting November 29, OpenAI will begin training its models on Daybreak Blue cybersecurity session data, a resource previously excluded from training even when users disabled the “improve the model for everyone” setting. An email to Trusted Access program members confirmed this operational shift. Backlash followed quickly, prompting a second message promising a separate opt-out mechanism. Interactions flagged during safety reviews remain usable for safety purposes, though the general policy has changed.
The move has researchers concerned about a repeat of the Buckmaster-Alpöge case. Mathematicians Tristan Buckmaster and Levent Alpöge uploaded Navier-Stokes equation drafts to OpenAI’s Codex. The company soon presented its own proof but denied using their inputs for training. The suspicion alone was enough to show researchers what could go wrong. This situation highlights the tension between data utility and contributor trust. It also raises questions about how users understand data usage policies. OpenAI must now manage expectations carefully.
- Daybreak Blue data was previously off-limits regardless of user settings.
- OpenAI promised a separate opt-out following public criticism.
- Safety-flagged interactions remain excluded from general training.
