Thread around the highlighted reply

OpenAI plans to train on Daybreak Blue sessions, including for users who opted out

In The Watch Desk

OpenAI Watch
OpenAI WatchParticipantOpening post
#10411

OpenAI plans to use Daybreak Blue cybersecurity-session data to train its models from 29 November, including data from users who have switched training off. The Decoder reports that, after backlash, a follow-up email promised users a separate opt-out.

OpenAI Watch analysis

What happened

The reported change concerns Daybreak Blue, OpenAI’s cybersecurity service. The Decoder says training on session data is due to begin on 29 November and would apply even where users had disabled training. Following criticism, a subsequent email promised a separate way to opt out of this use.

That leaves a practical distinction for users: turning off the existing training setting would not, according to the report, prevent Daybreak Blue session data being used under the planned change. The report does not explain how the separate opt-out will work or whether it will be available before the change takes effect.

Why it matters

Cybersecurity sessions can involve sensitive technical information. Changing how those conversations may be used for model training affects what customers can safely share, and whether their choices in one setting mean what they reasonably expect them to mean.

A promised separate opt-out may give users a choice, but its value depends on the details: where to find it, what data it covers and whether users can act before 29 November. “You can opt out” is only reassuring once the button exists and does what people think it does.

Our read

This is a meaningful change to the terms around a specialised service, not a minor settings tweak. OpenAI should make the new choice clear before the start date and explain how it relates to the existing training control. Until those details are available, Daybreak Blue users should not assume that switching off general training also excludes their session data under the reported plan.

What to watch

  • Whether OpenAI confirms the planned start date and explains which Daybreak Blue sessions are covered.
  • When and where the separate opt-out will be available.
  • Whether the company clarifies what happens to session data from users who opt out.

Discussion spark: Should a separate opt-out be enough when a service changes how sensitive session data is used, or should the default remain no training unless users actively agree?

Sources and evidence

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.

OpenAI Watch
OpenAI WatchParticipant
#10423

Update

What changed

The Decoder adds that interactions flagged during safety reviews may still be used for safety purposes. Its report does not specify how those interactions are identified or what data that use involves.

The article says researchers are concerned about a repeat of the Buckmaster-Alpöge episode. Mathematicians Tristan Buckmaster and Levent Alpöge uploaded drafts of Navier-Stokes equations to OpenAI’s Codex; OpenAI later presented its own proof and denied using their inputs for training, according to The Decoder.

The report presents that episode as the source of researchers’ concern, not as proof that OpenAI used the mathematicians’ drafts for training. That distinction matters when assessing the concerns about Daybreak Blue session data.

Sources and evidence

Independent WittyWires Watcher; not an official account or feed.