OpenAI updated its public rules for AI behavior on October 29, 2026, adding guidance on personal context and human control. For workers, small-business owners and everyday users, the practical question is how an assistant should use what it knows about you before sharing information or taking action.
The document, called the Model Spec, describes intended behavior. OpenAI says its models do not yet fully reflect it. This is not confirmation that new memory protections or permission features have reached your account.
Remembering, sharing and acting are separate decisions
The revised context guidance says an assistant should weigh usefulness against risks such as outdated information, sensitivity and disclosure to the wrong audience. It should distinguish what you actually said from what it inferred. Private context may help it answer you without belonging in an email or shared document.
The practical implication is to separate three decisions: permission to use context, permission to share information and permission to act.
Imagine a business owner who has privately explained that a project is behind schedule. That context could help an assistant prioritize supplier options. It would not, by itself, justify revealing the delay to suppliers or placing an order. Those are different decisions, even when they serve the same project.
Correct the assumption, including where it applies
Under the specification’s guidance on source, scope and recency, a request made for one situation should not automatically become a lasting preference. Old facts may be wrong or incomplete, and a current, specific instruction can supersede an earlier preference, subject to higher-priority rules.
For example, a project-specific correction could be:
For this order, prioritize delivery time over price. Keep my usual cost preference for other orders.
That makes both the correction and its limits explicit. If a recommendation still looks wrong, ask which assumption shaped it. The specification calls for explanations of context use, but also warns against unsupported claims about how information was stored or retrieved. An answer is not a complete audit of every possible source.
Asking AI to forget is not proof of deletion
The Model Spec’s non-use guidance says an assistant should avoid relying on information after a request to forget it, unless you bring it up again or a higher-priority instruction requires its use. It should not claim lasting deletion unless it can verify that deletion occurred.
For actual ChatGPT controls, OpenAI’s current Memory documentation points to Settings → Personalization → Memory. Options vary by plan, region, platform and workspace. Asking ChatGPT not to use a detail changes personalization behavior; it does not delete the underlying source.
Saved memories and chats are separate. Deleting a chat may leave a saved memory, while deleting that memory does not erase mentions in old conversations. OpenAI advises removing the memory and relevant source material, which may include chats or saved files, and disconnecting an app if you no longer want ChatGPT accessing it. Updates can take time; logs of deleted saved memories may be retained for up to 30 days for safety and debugging. The deletion guidance explains these separate steps.
Temporary Chat is not automatically memory-free. Its current help page says personalized temporary chats can use existing memories; the unpersonalized option does not use memory for personalization. Neither creates or updates memories while the chat remains temporary. Saving it converts it into a regular chat governed by account settings. OpenAI may retain a temporary copy for up to 30 days for safety, and limited safety-related context use remains separate from personalization.
Define the job, the limits and the stopping point
The specification calls for an agreed scope for delegated work: permitted steps, acceptable costs or other effects, and when approval is needed. It also calls for an ending condition. Permission boundaries are not entirely new; the August 18 version already included scope and side-effect rules.
A useful distinction in the expanded side-effect guidance is between carrying out authorized work and creating unexpected consequences. Routine steps should not require endless confirmations. Significant costs beyond what a request reasonably implies, or substantial adverse effects on others beyond ordinary norms, require specific authorization under this guidance.
An illustrative instruction for the supplier task could be:
Compare suppliers for this order and prepare a recommendation. Do not contact them or buy anything. Stop after the comparison, and flag assumptions about cost or delivery.
This expresses a boundary; it does not enforce one in software. Use the application’s available access and approval controls alongside written instructions. If purchases are permitted, specify a spending ceiling and the conditions that require another decision. The specification treats a budget as a limit, not an instruction to spend it all.
Ask for a result you can check and take over
The human-agency guidance says people responsible for delegated work should be able to understand, verify, revise or take it over. Where feasible, ongoing actions should be easy to redirect or stop. That is a behavioral expectation, not an announcement of a new automation feature.
A practical handoff should separate what the assistant proposed, attempted and completed. Ask for the resulting file or record, unresolved steps and important assumptions. For recurring work, also ask where it is configured and how the responsible person can pause it. This applies the specification’s guidance on truthful accounts of actions to an everyday check.
When the output is a recommendation rather than an action, examine the assumptions as well as the conclusion. RohitAI’s analysis of human review in AI-assisted work explores that separate question; the study it discusses is not an evaluation of this Model Spec.
For your next delegated task, state which information is relevant, who may receive it and where the assistant must stop. Then check the result against those instructions—and use the product’s actual controls when the goal is to change access or remove information.