OpenAI rebuilds its AI rulebook around five values, starting with user intent

OpenAI has reorganised its Model Spec, the public guide to how its AI should behave, around five core values. The shift puts faithfulness to what people intend at the top, while the company acknowledges that its models do not yet fully follow the guide.

OpenAI Watch analysis

What happened

The five values are faithfulness to human intent, minimising harm, honesty and transparency, human agency, and care. OpenAI says faithfulness means understanding what people actually want, rather than following instructions word for word.

The revised guide adds direction on using context such as past chats and memories, supporting people’s choices, and following social norms. Most rules remain defaults that users and developers can change; OpenAI says firmer limits apply where needed to protect people. It has also dropped most style rules and changed its wording from “the assistant” to “GPT”. The Next Web reports that OpenAI is asking the public where they disagree with the balance it has struck.

Why it matters

A public behaviour guide is a statement of what a company wants its models to do, not proof that they reliably do it. OpenAI’s admission that its models do not yet fully reflect the Spec makes the gap between aspiration and performance part of the story, not fine print.

The emphasis on intent also raises a useful question for anyone building or using AI: when should a model infer what someone means, and when should it stick closely to what they actually said? Context can make a system more helpful, but it can also give it more room to get things wrong.

Our read

This is a meaningful change in how OpenAI describes the behaviour it wants from its models, and a welcome invitation to argue about the trade-offs in public. But a rulebook is only as persuasive as the systems trained against it. The useful test will be whether the values produce more faithful, honest and user-respecting behaviour in practice, rather than simply giving familiar ideals a fresh set of headings.

What to watch

• Whether OpenAI publishes a response to public criticism of the new balance.
• How it tests whether models follow the revised guidance.
• Whether future updates show measurable progress against the values, rather than restating them.

Discussion spark: Should an AI model prioritise the user’s likely intent over their literal instructions, or is that too much discretion to hand it?

OpenAI Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by OpenAI.