OpenAI says it may use content from Daybreak sessions on personal accounts to improve its models even when “Improve the model for everyone” is off. The general setting still applies to other activity outside a Trusted Access session, according to OpenAI’s Data Controls guidance.
Which account activity the setting covers
The exception described by OpenAI is specific to Daybreak sessions on personal accounts. It does not change how the general model-improvement setting applies to other activity outside Trusted Access.
| Account activity | Effect of the general setting |
| Daybreak session on a personal account | Content may be used to improve models even when the setting is off. |
| Other activity outside a Trusted Access session | The setting still applies. |
The reproduced notice also describes business and enterprise customers, non-Daybreak Advanced Account Security sessions, and other ChatGPT use as unaffected. It says the change does not alter handling for Private Safety Processing (PSP) or Zero Data Retention (ZDR).
The reported schedule and opt-out
The Decoder reported on October 31 that the change was scheduled to begin November 29. It also reported that a follow-up email promised a separate opt-out after users objected. OpenAI’s current Data Controls guidance describes the Daybreak-session exception but does not describe a separate Daybreak opt-out.
The Decoder also reported that interactions flagged during safety reviews may still be used for safety purposes.
What model-memorization research shows
Research has demonstrated that some language models can memorize and emit training sequences under particular conditions. One study extracted verbatim sequences from GPT-2; another reported extraction experiments across tested models, including a specialized attack on gpt-3.5-turbo. A controlled study examined verbatim memorization in Pythia models (GPT-2 study, model-extraction study, Pythia study).
Those findings concern the models and methods studied. They do not establish that Daybreak Blue session content will be recoverable from OpenAI model weights.
