OpenAI's new rulebook says its AI should not just tell you what you want to hear
The guide behind ChatGPT's models is now built on five values, from honesty to care. OpenAI says its models do not fully follow it yet.
If you use ChatGPT, OpenAI just rewrote the rulebook for how its AI should act. On October 29, OpenAI updated its Model Spec. That is its public guide to how it wants its AI models to behave.
The new version is built around five values. They are faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. In plain words: understand what people really want, avoid harm, tell the truth, help people decide for themselves, and do work people can rely on.
One change stands out for everyday users. The AI should not pick a side just because it thinks that is what you want to hear. You can still ask it to take a certain point of view. It just should not guess that on its own.
Honesty also gets a clear meaning. The AI should be truthful about what it knows, what it does, and what it is not sure about.
The guide also tells the AI to use context, like past chats and saved memories, to understand what you want. OpenAI gives an example. Say you ask for help comparing job offers. The AI might recall that you care about predictable hours. But it should show that guess and make it easy to correct.
OpenAI is also dropping most of its detailed style rules. Style should now follow what each person prefers. The guide also now calls the model GPT.
There is a catch. OpenAI says its models do not fully follow the new guide yet. It is asking the public for feedback.
Open it and check the claim yourself. That is the point.