OpenAI plans to train models on Daybreak Blue cybersecurity session data that was previously off-limits
Reported by The Decoder1 source ↓
OpenAI will begin using Daybreak Blue session data for model training on November 29, including when users have disabled the general training setting. After backlash, it promised a separate opt-out.
Key points
- The data was previously excluded from training even when users had disabled the general training setting
- OpenAI promised a separate opt-out in a follow-up email
- Interactions flagged during safety reviews may still be used for safety purposes
Details
The Decoder reports that an email to Trusted Access program members announced the change. OpenAI plans to start using Daybreak Blue session data for model training on November 29, although users have turned off the “improve the model for everyone” setting.
The change prompted backlash, after which OpenAI sent a second email promising a separate opt-out. Data from interactions flagged during safety reviews can still be used for safety purposes. The Decoder also reports that researchers are concerned about a possible repeat of a case involving mathematicians Tristan Buckmaster and Levent Alpöge, whose uploaded drafts were followed by an OpenAI proof; OpenAI denied using their inputs for training.
Written from coverage as of