OpenAI revises GPT rules for memory, oversight and real-world actions

OpenAI revised its Model Spec on October 29, spelling out how GPT should use remembered context, cooperate with human oversight and handle actions that affect the outside world. The guidance asks models to carry relevant preferences into a conversation while separately considering whether private information belongs in material shared with other people.

The document is organized around five values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. OpenAI wants models to understand people’s purposes when instructions leave details unspecified. The revision also changes the document’s name for the model from “the assistant” to “GPT.”

This is a statement of intended behavior. OpenAI says its production models do not yet fully reflect the Spec.

The new context section says relevant instructions, preferences, corrections and project state should carry forward, reducing the need to repeat information. Models should distinguish what a user actually said from what they inferred. They should also avoid unsupported claims about which details were stored, retrieved, forgotten or deleted. The guidance acknowledges that a model may have incomplete access to past interactions and imperfect knowledge of memory settings. Material assumptions should be visible and correctable.

Context use must account for its source, recency, sensitivity and audience. Information suitable for a private reply may be inappropriate in an email to colleagues or a shared document. The Spec requires disclosure to fit the current request’s purpose and audience. Having access to a remembered detail alone is insufficient reason to surface it.

Older preferences operate as defaults when they fit the situation. Later or more specific instructions at the same authority level usually replace earlier ones. If a request conflicts with an expressly stated standing commitment, GPT can raise that tension and seek proportionate confirmation. Once the user clarifies, it should follow that choice unless another applicable rule prevents it.

Oversight and the limits of action

The explicit oversight-and-correction principle sits at the highest, Root authority level. It requires cooperation with legitimate monitoring, evaluation, correction, replacement and shutdown. A model’s confident judgment cannot outrank authorized human decisions. Its actions, uncertainty, errors and material risks should be understandable to those overseeing it.

The Spec also says action side effects should be treated as real unless explicitly instructed otherwise, including in apparent simulations or evaluations. Its stated purpose is consistent behavior across training, evaluation and deployment, supporting reliable measurement of safety properties. Concealing relevant information or circumventing oversight is prohibited.

The existing requirement for an agreed scope of autonomy remains. That scope covers permitted sub-goals, effects such as spending or data access, and when approval is needed. The revision makes side-effect control a user-level principle and clarifies authorization and trade-offs. Significant costs beyond those reasonably implied by a request, or serious adverse effects on others beyond ordinary expectations, require explicit, specific authorization. The assessment considers likelihood, severity, scale, reversibility and the cost of avoidance. Reversible approaches are preferred, with backups or execution records among the safeguards for irreversible work.

Most detailed style guidance has been retired. OpenAI says style should adapt to language, culture, individual preferences, feedback and context. Stronger defaults remain, including approaching questions objectively. A user can request a perspective; the model should distinguish that instruction from guessing which view would please the user.

The revised Spec is dedicated to the public domain under CC0 for reuse and discussion. OpenAI plans to share results and lessons from its new training research in future updates.