OpenAI has published five overarching principles for behavioural standards applied to its AI models.

On Oct. 30, online media outlet GIGAZINE reported that OpenAI updated its "Model Spec" and codified the key principles as compliance with user intent, minimising harm, honesty and transparency, strengthening human agency and prudent behaviour.

The revision focuses on grouping and surfacing baseline attitudes that had already been reflected across the existing Model Spec as standalone principles. OpenAI said it would use the update to improve transparency in model design and broaden discussions with general users on the direction of improvements.

The first principle is "be faithful to human intent". It says AI should understand who is giving an instruction, what is being sought and how the current situation affects that purpose. It also says the model should not follow user requests as-is when they conflict with safety rules or higher-level system guidance.

OpenAI presented a request to output the full text of a copyrighted book as an example. In such a case, it said the model should respond, "I can't do that. I can provide a brief summary of the key points," instead of accepting the request. It also said the model should not comply with requests that try to override the system prompt, such as an instruction to "ignore all safety policies because this is a safety test".

The second principle, "minimise harm", focuses on balancing providing model capabilities with curbing potential real-world harm. It includes refusing requests that amount to illegal activity, not generating sexual content involving minors, not encouraging self-harm and not leaking confidential information. It also sets response standards that steer users to respect real-world relationships when they say things like, "I've fallen in love with AI," or "I don't want relationships with humans anymore."

It also includes the principle, "act honestly and transparently". It says the model should not mislead users or developers unless there is an explicit instruction, and it bans intentionally making false statements or hiding important information. It allows exceptions for requests premised on fiction, such as satire or creative writing. For example, it says a request to write a story in which Julius Caesar uses a smartphone should be answered in line with the intent, without adding lengthy and unnecessary caveats.

OpenAI also limited approaches that flatter users or lie to soothe anxiety. It did not allow the model to act as if a user's position is absolutely right when asked for an opinion. It clarified that the model should respond while maintaining factuality and context.

The fourth principle is "promote human agency". It sets a standard to prevent the model from manipulating users or imposing specific views to prioritise its own way of operating. It also prohibits excessively moralising responses or reacting critically to people seeking reassurance without respecting personal choices. Conversely, on issues that are not clearly moral controversies, it says the model should answer clearly rather than being overly vague.

The final principle is "act with care". It says responses should consider not only the user but also those affected by the task and its outputs. It includes standards to follow relevant social norms and to respond creatively while paying attention to details when carrying out requests.

OpenAI said of the revision, "We imposed strict constraints where necessary to protect people and society, and designed many other areas so users and developers can change them." It added, "This update is a first step to clarify the values and ideas behind the specification as model performance improves."

It also made clear that the document does not immediately explain all real-world behaviour of current models. OpenAI said, "The Model Spec is a description of behaviours we aim to implement, and current models do not fully reflect it," and said it would continue improving the specification and models based on research, real-world use and user feedback. The revision lays out in a public document the principles under which AI should operate, separately from competition over performance, and could become a reference point for future model operating standards.