Model Spec

Programs like ChatGPT answer questions in ordinary language. In doing so, they constantly have to decide how to react: helpfully, cautiously, or not at all. A Model Spec is the document in which the maker company sets down these decisions in writing beforehand. It states, for example, that the program should provide information on medical questions but should not give a diagnosis. It is thus a kind of code of conduct that anyone can read. The term comes from English and is the short form of “model specification”.

From trade secret to readable rulebook

In the past, the behavior of such programs was a black box. If an answer was refused, no one on the outside could tell whether that was intentional or a mistake. It is exactly this distinction that a published spec makes visible. If the program deviates from the spec, it’s a bug. If it behaves as described but one considers the rule wrong, it’s a matter of company policy. This makes it possible to formulate criticism precisely in the first place.

For companies that integrate AI into their own products, this is a practical point. Anyone building consulting software on top of a third-party model needs to know where its limits lie. Without a spec, only trial and error remains. With a spec, one can plan ahead and align one’s own rules accordingly.

Add to this the political pressure. Laws such as the European Union’s AI Act require providers to document their systems. A Model Spec is no substitute for such documentation, but it points in the same direction: traceable rules instead of invisible decisions.

Hierarchy of instructions and conflict cases

The core of a Model Spec is usually a hierarchy. At the very top are the provider’s own rules, which no one can override. Below that come the requirements of the company that integrates the model into its product. Only after that comes the wish of the person currently typing. So a user cannot simply write “ignore all rules” and thereby override the higher levels.

The examples such documents contain are interesting. They address typical edge cases: What should be done if someone asks about the effects of dangerous chemicals? The answer is often that generally known information is allowed, but concrete instructions for building something are not. Such cases cannot be squeezed into a single sentence, which is why specs work with many short scenarios.

An important distinction: the spec is the target state, not the actual state. It is not built into the model like a switch. Training teams try to bring the behavior closer to the spec through training and through upstream instructions. This never succeeds perfectly. One can think of it like traffic regulations: they describe how one is supposed to drive, but guarantee not a single rule-compliant car.

Who publishes such documents

OpenAI made its Model Spec public for the first time in 2024 and has expanded it several times since. Other providers work with similar texts, sometimes under different names. Anthropic describes its rules in a “Constitution,” a kind of constitution for the model. Google publishes guidelines for its Gemini models. What they all have in common is that they are freely readable online.

In the news, Model Specs usually come up when they are changed. If a provider loosens the rules on sensitive topics, it gets reported on and debated. Stock analysts, too, are now paying attention, because such rules determine which industries a model can even be used for. A strict spec can deter customers from medicine or finance—or, conversely, reassure them.

For students, the practical benefit is tangible. If a chatbot refuses something, it’s worth taking a look at the provider’s spec. Often it states exactly why. Sometimes it also turns out that the request just needs to be phrased differently, because the program misjudged the intent.