OpenAI explained how Model Spec is structured — the code of conduct for its models

OpenAI has published a detailed description of its approach to Model Spec — an internal document that formalizes the values and behavioral rules of its models. This is the first attempt by the largest AI company to publicly explain the logic that guides GPT-4o, o3, and future systems in decision-making. Model Spec is not a technical document.

It is rather a constitution: a set of principles that establish priorities in situations where the goals of the user, operator, and OpenAI come into conflict. The hierarchy looks like this: at the top is broad safety — the model should not contribute to actions that could cause harm to humanity on a global scale. Second is broad ethics: adherence to moral norms, honesty, rejection of manipulation.

Third is alignment with OpenAI's principles. And only fourth is usefulness to a specific user. This means that in edge cases, usefulness loses.

If a request seems benign but carries systemic risk, the model should refuse. However, OpenAI directly acknowledges: hypercaution is as much a problem as recklessness. Models that refuse to answer legitimate questions, add unnecessary warnings, or adopt a moralizing stance undermine trust and product value.

The company clearly signals: the goal is balance, not maximum tightening of restrictions. The document describes a three-level trust system. OpenAI sets basic rules that cannot be circumvented.

Operators — companies and developers using the API — gain the right to adjust the model's behavior within these limits: they can permit content blocked by default or add additional restrictions for their audience. Users act within the boundaries set by the operator. The more context and verified trust — the more flexibility.

A separate topic in the document is the question of internal states of the model. OpenAI does not claim that its models possess consciousness or subjective experience, but neither does it entirely reject this possibility. Model Spec states that models can have something like functional emotions — not in the sense of real feelings, but as internal states that affect behavior.

The company commits to seriously studying this question, not closing it off with the statement "this is just text statistics." In conditions of increasing regulatory pressure — EU AI Act, executive orders in the US, initiatives in the UK and Japan — companies able to explain the logic of their systems gain a competitive advantage. Model Spec is a tool of accountability.

When a regulator asks "why did your model do X," OpenAI now has a public framework to answer. When a corporate client demands predictable behavior — the document provides a basis for that conversation. Model Spec also changes industry standards.

Anthropic published Constitutional AI and safety principles. Google DeepMind regularly releases safety papers. Now OpenAI makes its approach equally transparent — and this becomes an unspoken requirement for any serious player.

Companies without such a document increasingly look like less reliable partners. For users, Model Spec means one thing: the behavior of ChatGPT and API models becomes more predictable. If something works unexpectedly — you can refer to the document and understand which principle took effect.

This is not a guarantee of perfect answers, but it is honesty about what compromises are built into the system.