Evidence at a glance
A Spec for Situations Rules Cannot Fully Anticipate
On October 29, 2026, OpenAI updated the Model Spec, its public description of how its models should behave. The revision addresses a growing challenge: as models become more capable, they may take actions with greater consequences, while user instructions do not always make their background, purpose, and limits explicit. OpenAI argues that models must do more than interpret the words in front of them. When instructions leave gaps, they should judge human intent and values while remaining subject to human oversight.
The change matters not because it adds five memorable terms, but because it moves the governance problem into less specified situations. A behavior guide can list rules and defaults, but a model facing a case those rules do not anticipate also needs guidance on how to judge—and on how that judgment remains faithful to people. Rules still matter. What changes is that the room for judgment beyond explicit rules becomes more consequential and harder to demonstrate as reliable.
Five Values, Not Five Standalone Commands
The updated Spec organizes its guidance around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care for those affected. These are not five independent checks. Faithfulness means helping people pursue their aims, while harm minimization requires considering foreseeable consequences. Supporting agency means helping people understand and choose rather than quietly choosing for them. Honesty and care also require clarity about what the model knows and attention to the circumstances of affected people.
The framework therefore asks more than whether a model obeys a user. It addresses cases where an immediate request, the user’s longer-term interests, and the effects on others may not align. OpenAI’s example is a model comparing job offers by drawing on a previously stated preference for predictable hours, while making that assumption visible and easy to correct. The example clarifies the intended direction, but also the difficulty: an old preference may help interpret intent without necessarily remaining valid.
Hard Boundaries and Defaults Have Different Jobs
The values do not give a model unrestricted discretion. OpenAI describes a layered authority structure: OpenAI sets boundaries and trains models to respect them, developers direct applications within those boundaries, and users direct the model’s work subject to those constraints. The update also revises instruction priority and conflict handling. It retains necessary hard safety boundaries while treating many other behaviors as defaults that users or developers can override.
That distinction is useful in application design. Hard boundaries can carry constraints that an individual application or user should not be able to remove at will, while configurable defaults leave room for product, preference, and workflow differences. But making a default overridable does not settle conflicts by itself. If a model misjudges who can override what, or treats contextual hints as current instructions, its behavior can still depart from expectations. The announcement says priority and conflict rules have changed, but does not provide operational detail sufficient to reproduce those judgments.
Context Makes Intent Inference More Useful—and More Fragile
The revision more explicitly asks models to consider context, including information from conversations and memory, when inferring intent. For tasks that are easy to misread from a single sentence, this can connect earlier and later statements and reduce mechanical execution that misses the user’s actual goal. But the more a model relies on history to infer preferences, the more an outdated preference, mistaken memory, or incomplete background can shape its actions.
That makes faithfulness a problem of correction and control as well as response quality. OpenAI’s job-offer example emphasizes exposing assumptions so users can readily correct them, rather than burying inferences in an answer. Yet the public materials do not explain how users can inspect or clear memories used in a judgment, or set a general standard for handling mistaken intent inference. For applications involving consequential decisions, designers should not equate using context with understanding a person better. They should make important assumptions visible and preserve opportunities to confirm them.
The Commitments Still Need Behavioral Evidence
OpenAI says it is exploring new training methods, but has not disclosed the specific process or experimental results. It has published no rate of adherence, evaluation scores, or measured improvement, and it acknowledges that current models do not yet fully reflect the Spec. The update therefore demonstrates a change in governance goals and public guidance. It does not demonstrate that models consistently understand intent, resolve value conflicts well, or behave consistently across contexts.
For technical leaders, a practical starting point is to separate non-overridable safety constraints from configurable behavior, then make the context and assumptions a model is expected to infer explicit. Evaluation should go beyond checking literal compliance. It should include ambiguous intent, obsolete preferences, conflicts between a request and the interests of others, and actions with side effects, while checking whether users can notice and correct errors. The five values offer a shared language for designing such tests. Until public validation is available, however, they remain design commitments, not a certificate of reliability.