OpenAI explains how the Model Spec works, the code of conduct for its AI models

OpenAI has published an extended description of its approach to the Model Spec, an internal document that formalizes the values and behavioral rules of its models. It is the first attempt by the largest AI company to publicly explain the logic that GPT-4o, o3 and future systems follow when making decisions.

The Model Spec is not a technical document. It is closer to a constitution: a set of principles that set priorities when the goals of the user, the operator and OpenAI come into conflict.

The hierarchy looks like this. First comes broad safety: the model must not facilitate actions that could cause harm to humanity on a global scale. Second is broad ethics: following moral norms, honesty, and refusing to manipulate. Third is compliance with OpenAI's principles. Only in fourth place comes helpfulness to the individual user.

This means that in borderline cases, helpfulness loses. If a request seems harmless but carries systemic risk, the model should refuse.

However, OpenAI states plainly that excessive caution is as much a problem as recklessness. Models that refuse to answer legitimate questions, add unnecessary warnings or take a moralizing stance erode trust and the value of the product. The company clearly signals that the goal is balance, not maximum tightening of restrictions.

The document describes a three-tier trust system. OpenAI sets baseline rules that cannot be bypassed. Operators, meaning companies and developers who use the API, may customize the model's behavior within those bounds: they can allow content that is blocked by default, or add extra restrictions for their audience. Users act within the limits the operator sets. The more context and verified trust, the more flexibility.

A separate topic in the document is the question of the model's internal states. OpenAI does not claim that its models have consciousness or subjective experience, but it does not fully rule out the possibility either. The Model Spec says that models may have something like functional emotions: not in the sense of real feelings, but as internal states that influence behavior. The company commits to studying the question seriously rather than closing it with the statement "it's just text statistics."

Under growing regulatory pressure, including the EU AI Act, executive orders in the US, and initiatives in the UK and Japan, companies that can explain the logic of their systems gain a competitive advantage. The Model Spec is an accountability tool. When a regulator asks "why did your model do X," OpenAI now has a public framework for answering. When a corporate customer demands predictable behavior, the document provides a basis for that conversation.

The Model Spec also shifts industry standards. Anthropic has published Constitutional AI and its safety principles. Google DeepMind regularly releases safety papers. Now OpenAI is making its approach just as transparent, and this is becoming an unspoken requirement for any serious player. Companies without such a document increasingly look like less reliable partners.

For users, the Model Spec means one thing: the behavior of ChatGPT and the API models becomes more predictable. If something works unexpectedly, you can consult the document and work out which principle kicked in. It is not a guarantee of perfect answers, but it is honesty about the trade-offs built into the system.