OpenAI rebuilds its Model Spec around five values

Faithfulness comes first. Under it, a model reads each instruction for the goal behind it, weighing what the user, the developer and OpenAI itself are after and care about. The remaining values have the model limit harm, admit the limits of what it knows and leave people’s choices to them. By OpenAI’s own account, today’s models still fall short of the new text in part, and it is trying new training methods to bring them in line, Reuters reported in a story carried by Allwork.Space. Reuters sets the approach beside Anthropic’s, which guides Claude with a written constitution.

Our rights register, read on 24 September 2026, holds OpenAI’s rules for what you do with ChatGPT’s images. The output is yours. OpenAI’s Sharing & Publication Policy, dated 14 November 2022, asks that content posted to social media be labelled as AI-generated so plainly that no reader could overlook it, and the terms bar passing output off as human-made. The policy does not say whether a paid ad counts as a post.

The new spec governs how honest the model is inside the chat; whether an ad built from its images says it is AI-made still rests with the advertiser, under a 2022 policy that leaves paid ads unaddressed.