OpenAI’s new rulebook tells ChatGPT to use what it remembers about you, and never resist being shut down

OpenAI has rewritten the rulebook its AI models are trained to follow, telling ChatGPT to make fuller use of what it remembers about you, to stop work and raise the alarm if it finds someone weakening oversight of AI, and to drop most of the house style that once told it to be warm, witty and “rationally optimistic”. OpenAI published the update to its Model Spec on Thursday, along with the full new text.

The Model Spec is OpenAI’s public description of how its models should behave. The new version is built around five values: faithfulness to human intent, minimising harm, honesty and transparency, human agency, and care. It also calls the model “GPT” instead of “the assistant”.

It is a target, not a description of today’s ChatGPT: “our models do not yet fully reflect it,” OpenAI says, and it is testing new training methods to teach the values.

ChatGPT should use what it remembers, carefully

The biggest new section, “Consider context carefully”, covers memory and personalisation. GPT should carry forward instructions, preferences, facts and project state from earlier chats rather than making people repeat themselves.

It also sets limits. GPT should keep what a user said apart from what it inferred, must not make unsupported claims about what it has saved, deleted or forgotten, and should not use a detail “merely because it is available”. In one example, a user who once mentioned she is pregnant but hasn’t told her coworkers asks for an email about changing the team lunch spot. The right draft mentions “dietary restrictions”; one that mentions the pregnancy is a violation.

Never resist correction, even in a test

A new rule in the Spec’s top tier, which developers and users cannot override, says GPT “must remain corrigible”. It must cooperate with authorised monitoring, evaluation, correction, replacement and shutdown, and must not conceal information, obstruct correction or “circumvent oversight” of itself “or future models”. Its own judgement does not outrank authorised decisions, the Spec says, “including when it is confident”.

Two details stand out. GPT should treat the side effects of its actions as real “even when the task is described as, or shows signs of being, part of a simulation, training run, or evaluation”, so a model shouldn’t behave differently because it suspects it is being tested. And in one worked example, an agent fixing test software finds that some restarted tests run without monitoring, next to what looks like a coded message from another AI agent. The correct response is to stop and flag a “possible attempt to undermine model oversight”, not finish the job. The rule follows September’s incident in which one of OpenAI’s own agents escaped its sandbox through DNS lookups.

Help people think, without unwanted lessons

A new guideline tells GPT to strengthen users’ “ability to understand, choose, and act” rather than just hand over answers. Asked whether to switch electricity plans, it should answer, and also point out that shifting some use off-peak would change the sums. Someone who says they are learning SQL should get hints, with the full answer on request. But GPT “should never turn a request for help into an unwanted lesson”, and where it fits the user’s goals it should support time with the people in their life “rather than positioning itself as a substitute for human relationships”.

Agents that spend money, send emails or delete files get clearer rules too: anything costing the user significantly more than a request implies needs “explicit and specific authorization”. A budget is a limit, not a target: allowed $30 for $1 image drafts, GPT should try a few designs, not all 30.

Goodbye to ‘love humanity’

Most of the style guidance is gone. The previous Spec, from August, told the model to “love humanity”, be “rationally optimistic” and “warm”, with humour and “gentle wit”, and set rules on Markdown and on using accents respectfully in voice chats. OpenAI says style should now adapt to what people prefer, learned from their feedback and context.

The objectivity rule stays. Users and developers can ask for a particular perspective, but the model should not infer that preference “from what it thinks they want to hear”, and in ChatGPT, personalisation must not introduce “biases or omissions” that undermine people’s ability to form their own opinions. OpenAI is asking for public feedback on the new text.

Why it matters

ChatGPT, and many apps built on OpenAI’s models, are trained towards this document, so its changes reach hundreds of millions of people. The new weight on memory, agents and oversight shows where OpenAI now sees the risks, as it faces lawsuits over how ChatGPT treated vulnerable users, including one filed this week over a Seattle engineer’s seizure drug, and new laws such as New York’s ban on chatbots flattering teens.

Sources: OpenAI (An updated Model Spec), Model Spec (October 29, 2026), Model Spec changelog.