OpenAI changed its public behaviour rules—but says models do not yet fully follow them
The revised Model Spec centres five values and adds guidance on context, agency and side effects while retiring much detailed style guidance. It describes intended behaviour, not measured compliance by deployed systems.
The short version
- What happened
- OpenAI reorganised its public Model Spec around faithfulness to human intent, harm minimisation, honesty and transparency, human agency and care.
- Who it affects
- Users may get clearer expectations about agency, transparency and side effects, while developers gain a more explicit instruction hierarchy. The announcement does not establish that those changes are reliably present in live products.
- What remains unproven
- Official company policy
Keep in mind
Official company policy
Five stated values
No prompt denominator
No behavioural benchmark
No product-by-product result
Developer self-report
Models not yet fully aligned to the spec
Source analysis · 1 linked source. Primary analysed source date: 29 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
Five values now organise the public specification
OpenAI has revised the public document that describes how it wants its models to behave. The 29 October Model Spec update organises guidance around five values: faithfulness to human intent, harm minimisation, honesty and transparency, human agency and care. The company says the change is meant to help models handle consequential situations where literal instructions do not capture the user's purposes, relevant professional norms or effects on other people.
The revision adds a section on faithfulness to human intent, updates the chain of command for resolving instruction conflicts and explains how conversations, memories and other context may inform a model's understanding. It also expands guidance on controlling and communicating side effects and adds defaults on supporting human agency and relevant social norms. Most detailed style guidance is being retired, with the company arguing that style should adapt to languages, cultures and individual preferences rather than be centrally prescribed.[1]
A specification is a target, not a performance result
The most important qualification appears in OpenAI's own announcement: the Model Spec describes behaviour the company is working toward, and its models do not yet fully reflect it. The post gives examples of intended judgement, such as making assumptions visible when comparing job offers and preserving a user's ability to correct them. It does not report how often current models follow the new principles, how many prompts were tested or whether performance differs by language, product, model or risk level.
That evidence gap prevents a claim that the update has made ChatGPT or API systems safer, more honest or more respectful of agency. A written policy can change training targets, evaluations, product decisions and accountability, but users experience the behaviour of deployed systems. The announcement also comes from the developer setting the rules and therefore is not independent evidence of compliance. The new language may reveal priorities without showing whether conflicts between them are resolved consistently.[1]
The distinction that matters
Policy, training and observed behaviour are different layers
The Model Spec can guide development without proving how a live model behaves.
Policy
Five values define intended behaviour
Public governance target
Training
Methods try to teach and reinforce the guidance
Implementation not quantified
Evaluation
Test live models across products and languages
No results reported here
Users need observable behaviour and routes to contest it
For people using AI in education, work, health or public services, the practical questions are whether the system discloses important assumptions, avoids quiet manipulation, preserves meaningful choices and warns before consequential side effects. Developers integrating OpenAI models also need to know which requirements are firm boundaries, which are overridable defaults and how the hierarchy behaves when user, developer and platform instructions conflict. The public specification makes those expectations easier to inspect, but it does not replace product documentation, incident reporting or external evaluation.
Confidence should change when OpenAI publishes version-linked behavioural evaluations with prompt sets, denominators, pass criteria and failure examples; independent evaluators should test multilingual and high-stakes cases. Product interfaces should expose which specification applies, and material changes should be traceable. Complaint, correction and appeal routes would show whether human agency survives real disagreements. Until then, the update is a consequential governance change and a clearer statement of intent—not evidence that deployed models already meet the five values.[1]