OpenAI just published the playbook for how its models decide what to say and what to refuse.

The Summary

  • OpenAI released detailed documentation on its Model Spec, the framework governing how ChatGPT and other models behave in the wild
  • This is the constitution for AI decision-making: what gets answered, what gets blocked, and who gets to decide
  • The real story is accountability architecture. As models gain agency, the rules hardwiring their behavior become infrastructure, not just policy.

The Signal

The Model Spec is OpenAI's attempt to make explicit what has been implicit: the decision tree that determines when a model helps you and when it lectures you. It covers refusals (when models say no), style defaults (how formal or casual responses skew), and what OpenAI calls "conflicting objectives," the impossible balance between being helpful and being safe.

This matters because models are becoming agents. When your AI just answered questions, vague guidelines were fine. When it's booking flights, managing your calendar, and negotiating on your behalf, the governing rules become critical infrastructure. The Model Spec is OpenAI saying: here's what we hardwired, here's why, and here's how we'll update it.

The framework prioritizes developer and user intent over OpenAI's preferences, except when safety is at stake. That exception is doing heavy lifting. In practice, it means if you ask a model to be more direct or skip the safety preambles, it should comply, unless the request trips hardcoded safety rails. The Spec also introduces explicit rules about following laws (models default to obeying the law of the user's jurisdiction) and respecting social norms (models adapt based on context).

What's most telling is the public commitment to iteration. OpenAI positions this as a living document, updated as models evolve and edge cases emerge. That's critical, because today's polite refusal to help with a legal but controversial request might look absurd in 18 months when competitor models have no such guardrails.

The Implication

If you're building on OpenAI's APIs, the Model Spec is now your compliance map. Know where the rails are. If you're in the agent economy, watch how other model providers respond. This is OpenAI staking out territory on values alignment, and the industry will either converge or fragment around competing specs. For users, this is a preview of the governance layer that will sit underneath the agent economy. You want models that follow your intent. You also want to know where they won't. The Spec makes both legible.

Source: OpenAI Blog