NewsPublished:
OpenAI Rebuilds Its Model Spec Around Five Core Values and Drops Most Style Rules
The updated Model Spec centers on intent, harm, honesty, agency and care, adds context guidance, retires style rules and calls the model GPT.
OpenAI has published a big revision of its Model Spec, the public document that describes how it wants its models to behave. The headline change isn't a new rule so much as a new structure: the whole thing is now organized around five core values instead of a long list of separate guidelines.
The five values
According to OpenAI, the updated Spec is built on:
- Faithfulness to human intent, meaning understanding what people actually want rather than just following instructions literally
- Minimizing harm, with firm safety boundaries and attention to foreseeable consequences
- Honesty and transparency about what the model knows, does and isn't sure about
- Human agency, strengthening people's ability to understand and decide for themselves
- Care for the people affected and for doing work others can rely on
OpenAI says most of the guidance is still written as defaults that developers and users can change, with hard limits kept only where it thinks people need protecting.
What actually changes
There are new sections on reading context carefully, including how conversations and memories should shape what the model assumes you mean, plus new defaults on supporting human agency and following social norms. The guidance on side effects, meaning the knock-on results of what a model does, has been expanded too.
The part I find most interesting is what got removed. OpenAI is retiring most of the detailed style rules, saying style should adapt to what people prefer. One stronger default stays: the model should approach questions objectively and not lean toward a viewpoint just because it guesses that's what you want to hear. The Spec also now calls the model "GPT" instead of "the assistant".
My take
For anyone who uses these tools to brainstorm fan theories, draft translations or plan cosplay builds, the anti-flattery default matters more than any of the philosophy. OpenAI is upfront that its models don't fully follow the Spec yet, and it's asking for public feedback. You can read the details in OpenAI's announcement.
Source: openai.com