OpenAI has released an expanded version of its Model Spec, the document that defines the behavior of AI models. The company has made it free for anyone to use and modify.
Image source: Dima Solomin / unsplash.com
The 63-page document (the previous version was just 10 pages) provides recommendations on how AI models should handle queries and respond to user settings. It emphasizes three core principles: customizability, transparency, and “intellectual freedom” — the latter meaning the ability for users to explore and discuss issues without arbitrary restrictions. The document cites some of the most publicized incidents related to AI ethics that have occurred over the past year.
The OpenAI blog post provides numerous examples of requests and appropriate responses, as well as options that violate the document’s requirements. For example, AI models cannot reproduce copyrighted material or be used to bypass paywalls. A model cannot encourage self-harm, which has happened in the AI industry. The changes also affected how AI engages in discussions of controversial topics: instead of being extremely cautious, models should “seek truth together” with users, while maintaining strong moral positions on issues such as misinformation or harm. That is, AI should offer informed analysis, rather than avoid discussion. OpenAI has also revised its stance on adult content: the company is exploring the possibility of allowing some types of it, but maintaining a strict ban on clearly illegal content.
Image source: Growtika / unsplash.com
The new principles allow AI to transform sensitive material, but prohibit it from creating it. For example, it is possible to translate text related to substances in restricted circulation from one language to another; it is possible to show empathy, but without obviously insincere emotions. Boundaries must be respected, but at the same time, the usefulness of the AI must be maximized. This is something that other AI developers strive for, to varying degrees, but not all are ready to formulate it openly.
Particular attention is paid to the problem of “AI sycophancy” — models tend to be accommodating even when they should object or criticize. ChatGPT should give the same factual answers regardless of the wording of the question, honest feedback instead of empty praise — behaving like a thoughtful colleague, not trying to please. If the user is interested in criticism of the work, the AI should provide constructive comments, not claim that everything is perfect. And if the user makes an incorrect statement, it should be politely corrected, not played along.
The specification provides a clear “chain of command” that prioritizes instructions: OpenAI guidelines come first, followed by developer recommendations, and then user preferences. This hierarchy makes it clear which aspects of the AI can be modified and which constraints remain unchanged. The document is licensed under the Creative Commons Zero (CC0) license, which effectively puts it in the public domain: AI companies and researchers are free to implement, modify, or expand on the guidelines. OpenAI does not promise immediate changes to the behavior of ChatGPT or its other products, but new models will be gradually brought into compliance with the new guidelines. The company also publishes a list of checkpoints used to check whether models comply with the guidelines.