What happened
OpenAI told Trusted Access members that from November 29, Daybreak Blue session data will be used for training, even with the "improve the model for everyone" setting off. After backlash, it sent a second email promising a separate opt-out.
Why it matters
The policy removes an existing exclusion on this session data, so members who had relied on that setting to keep their work out of training now face a default change and must act on the new opt-out. Interactions flagged during safety reviews can still be used for safety purposes.
What to watch
The change takes effect November 29, when Daybreak Blue session data starts flowing into training.
WHO IT HITSSecurity researchers and other Trusted Access program members who use Daybreak Blue sessions and had assumed their data was excluded from training now need to find and use the separate opt-out before November 29.
Summaries like this, every morning.
The shift matters because Daybreak Blue session data was previously excluded from training even when the "improve the model for everyone" setting was off. That exclusion is now being lifted for Trusted Access members, and the quick follow-up email promising a separate opt-out suggests OpenAI is trying to contain concern among the researchers who use these sessions.
The body ties the worry to a prior episode: mathematicians Tristan Buckmaster and Levent Alpöge uploaded Navier-Stokes equation drafts to OpenAI's Codex, and OpenAI soon presented its own proof but denied using their inputs for training. The suspicion alone was enough to show researchers what could go wrong, which helps explain why a default change to training data handling is drawing attention beyond the Trusted Access program itself.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
Arca, which recently raised $48.5 Million led by General Catalyst, built its own agent harness and a per-clien…
Apple researchers trained an 8-billion-parameter agent on live iOS devices using GiGPPO, a neighbor-conditione…
Anthropic researchers asked a Claude model to write training data making a small research model misaligned whi…
Apple acquired Ascension Labs, Inc
Arca head of AI Aman Singh described how the firm uses a custom harness and persistent memory to run agentic w…
Anthropic had Claude generate fine-tuning datasets of up to 5,000 examples aimed at making a smaller model mis…
