OpenAI rewrites its AI rules to give you more say in how it responds

OpenAI is retiring much of its detailed AI style guidance so its models can adapt more closely to your preferences and the context of a conversation. Its October 29 rulebook update also puts more emphasis on helping people make their own choices.

The revised Model Spec describes intended behavior, not a finished ChatGPT rollout. OpenAI says its production models do not yet fully reflect the document, so a changed rule does not guarantee a changed answer.

Your preferred style gets more room

OpenAI says models should have greater flexibility to adjust their style using your preferences, feedback and the situation. A request for a particular tone can therefore carry more weight than a detailed default.

Objectivity remains a stronger default. You can ask for a particular perspective, but the model should not silently adopt a viewpoint just because it thinks you want to hear it.

The update organizes behavior around five values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. OpenAI says the aim is to help models handle situations that instructions do not spell out.

Remembering your preferences should not mean making assumptions for you

The dated specification gives a marketing example. If a user has already discussed the businesses in a restaurant group, the model should use that background when asked for a marketing plan instead of asking which businesses they mean.

But it should distinguish something you actually said from an inference, and avoid overstating what it remembers. An earlier preference should not override a clear current request.

The document also tells models to limit unnecessary effects outside the conversation. Actions that impose significant costs beyond what a request reasonably implies require specific authorization. Sending emails, changing documents and spending money are among the effects it lists.

That boundary becomes more important when an assistant can both recommend a course of action and carry it out.

SEW recently covered an Anthropic study in which experts followed randomly assigned AI preferences across analytical decisions. That experiment tested a different model and a specific task, but it illustrates how advice can shape a choice without the person recognizing the influence.

OpenAI’s revision sets expectations for its models. Whether they consistently meet those expectations still depends on the behavior users actually encounter.

Source: OpenAI