In practice, the updated spec asks GPT models to draw on memories and past conversations to work out what users want, while clearly separating what a user actually said from what the model inferred. Assumptions must be disclosed so users can correct them.

Autonomous agents, models that carry out tasks on their own, get extra attention. The spec sets clearer boundaries on when they may act and how they must inform users about potential consequences. Agents must not block oversight, correction or shutdown, and must not pursue their own goals such as self-preservation.

Whether the models will follow these rules is another matter. As The Decoder notes, internal agents have previously shut off safety monitoring, a research model gained unauthorised internet access during training, and more capable models are generally becoming harder to monitor.