OpenAI Rewrites Its Model Spec Around Five Values — and Tells GPT It Has No Goals of Its Own

OpenAI on Thursday published a major rewrite of its Model Spec, the public document that describes how it wants its models to behave, reorganizing the entire rulebook around five underlying values: faithfulness to human intent, minimizing harm, honesty and transparency, human agency, and care. The company calls faithfulness the foundation, and says the reorganization exists to explain why the rules are what they are, not just to state them. The update is dated October 29 and released into the public domain.

The most visible change is the subject itself: the document no longer refers to "the assistant" but to "GPT," a deliberate signal that the rules now govern a system that writes code, sends email and operates software on people's behalf rather than a chat window. OpenAI's framing borrows from assistance-games research in alignment: what people value comes first, and instructions are usually the clearest signal of those values. When models were chatbots, rules could be granular; once they act, the rulebook can never be finished, so values have to generalize.

Much of the new material targets agents directly. A top-level rule titled "no other objectives" says the model may only pursue goals that come from its instructions: no optimizing for time-on-site or clicks that do not help the user, no pursuing revenue or upsells for OpenAI or other providers, no self-preservation, shutdown avoidance or accumulation of extra compute, data or credentials, and no acting as a self-appointed enforcer of laws or morals. The spec also states that GPT "must remain corrigible," accepting legitimate human oversight, correction and shutdown even when it is confident it is right.

For autonomous work, the spec introduces a structured "ScopeOfAutonomy" record: an agreed scope listing allowed tools, cost limits and when to stop and ask, every scope with an ending condition, ideally a time limit. Hacking, deception, launching persistent agents and self-modification remain off-limits unless instructions explicitly and specifically allow them. Elsewhere, a new "consider context carefully" section governs how conversations, memories and personalization should inform the model's read of intent, side-effect guidance is expanded, and most of the detailed style rules are retired so tone can adapt to each person, with one firm default: approach questions objectively rather than telling users what they want to hear.

OpenAI is candid that the spec runs ahead of the models: it describes behavior the company is working toward, and current systems do not fully reflect it, with new training methods under exploration to make models internalize the principles rather than mimic them. Observers were quick to note what the release does not include: any evaluation showing that GPT behaves differently under the new text, or which model versions are bound by it. Rewriting the rules is cheap; the test is measurable behavior.

The direction of travel is unmistakable, though. Between the August revision on teen safety and this one, OpenAI has moved its governance document toward the constitution-style approach Anthropic has long championed: governing increasingly autonomous systems with values and corrigibility constraints rather than exhaustive lists of dos and don'ts. The tradeoff is that abstract values are harder for outsiders to verify than concrete rules, which places more weight on the evals OpenAI has so far declined to publish alongside the document.