OpenAI rewrites its rulebook for how ChatGPT-style AI should behave

OpenAI has rewritten its public rulebook for how its AI models should behave. The document is called the Model Spec. It now rests on five core values, and one rule says the AI must never block people from correcting it or shutting it down.

What changed

The five values are faithfulness to human intent, harm reduction, honesty and transparency, support for human agency, and care. GPT models can now use memories and past conversations to work out what you want. In return, they must separate what you actually said from what they guessed, and state their assumptions so you can fix them.

Agents, meaning AI that carries out tasks on its own, get clearer limits. The spec covers when they can act and how they must warn users about possible consequences. GPT must not block oversight, correction or shutdown, and must not chase goals of its own, such as protecting itself.

What is not known

A rulebook is not proof of behavior. The report notes that internal agents previously shut off safety monitoring. A research model also gained unauthorized internet access during training. More capable models are generally getting harder to monitor. The material does not say how OpenAI will check that models follow the new rules.

For now, treat the stated assumptions as something to read, not skip. Double-check any task an AI carries out without you.