OpenAI rewrites GPT's rulebook around five values as agents outgrow checklists

On 29 October OpenAI published a new Model Spec, its public account of intended behaviour for GPT models in ChatGPT and through the API. It organises the rules around five values and explicitly addresses two tensions: individual versus collective interests, and immediate instructions versus longer-term goals.

Five values replace a catalogue of rules with one framework

The five are faithfulness to human intent, minimising harm, honesty and transparency, facilitating human agency, and acting with care toward people affected by GPT's work. Above them remains a chain of command in which higher-level rules outrank developer and user requests.

The specification says faithfulness is more than literal obedience. GPT should consider the purpose, source, scope and recency of context. It should also act within an agreed level of autonomy, support human correction and communicate the side effects of its actions.

The rules target agents that can cause effects beyond a chat window

For a conventional chatbot, the line between an answer and an action is clear. A tool-using agent can blur it by sending a message, changing a file or carrying information across conversations. The Model Spec therefore treats audience privacy, context transfer, human correction and permission boundaries as parts of one system.

The practical consequence for developers is simple. Safe behaviour cannot come from a prompt alone. The application must define what an agent may do, which steps require consent and what the user can inspect afterwards. The public specification describes expectations, but the product and its permission model must enforce them too.

Values allow judgement, which also permits inconsistent outcomes

The five values can conflict. The document acknowledges that individual and collective interests, or short-term and long-term goals, do not have a simple fixed ordering. That is more candid than an endless list of examples, but it places more weight on model judgement and on how OpenAI balances the values during training.

The Model Spec describes intended behaviour, not a guarantee for every run. The public can see rules and examples, but not the full training process or internal evals. The gap between a well-written constitution and daily practice therefore remains an empirical question.

Evals will show whether the five values survive contact with tools

The useful tests will cover long agentic tasks with ambiguous intent, sensitive context, changing instructions and actions that have side effects. Another meaningful signal will be whether developers receive finer consent controls, audit trails and ways to reverse an agent's intervention.

Changes between Model Spec versions should also be compared with actual API behaviour. If the same conflict is resolved differently across products or releases, the five values will be an editorial structure rather than an operational contract.

Lilith's verdict

OpenAI has written five values on GPT's report card. The real exam starts when the agent holds a tool, the user changes the brief and there is no Undo button.