Key insights

• The Model Spec is written for humans first, models second. It is a social contract and communication tool, not a training script.
• Honesty now explicitly outranks confidentiality after OpenAI discovered that keeping developer instructions secret could push models toward covert, deceptive behavior.
• Reasoning models follow the spec better because deliberative alignment teaches them to reason through policies in their chain of thought, not just match patterns.
• The three top-level goals in the spec parallel Asimov's robot laws, but deliberately avoid a strict hierarchy because rigid priority lists break down at the edges.

In Brief

OpenAI has a public document called the Model Spec that explains how its AI models are supposed to behave. It is about 100 pages long, covers everything from honesty and safety to tone and personality, and is open source on GitHub so anyone can read or fork it. In this episode of the OpenAI podcast, host Andrew Mayne talks with Jason Wolfe, a researcher on OpenAI's alignment team who helped write and now maintains the spec. They discuss how the spec works in practice, what happens when rules conflict, and how the document has changed over time based on real-world feedback and incidents.

What is the Model Spec?

The Model Spec is OpenAI's attempt to write down, in plain language, how it wants its AI models to behave. Think of it as a company handbook, but for an AI instead of a human employee.

A few things the spec is not: it is not a claim that models already follow it perfectly. It is not a technical training script that feeds directly into how models are built. And it is not a complete description of everything that happens when you use ChatGPT (there are separate safety systems, product features like memory, and usage policies that sit outside the spec). What it is, first and foremost, is a communication tool for people. OpenAI wants employees, developers, users, and policymakers to be able to read it and understand what the company is aiming for.

As Wolfe puts it, the spec is a north star, pointing toward where OpenAI is trying to go. The actual models may not always be there yet.

---

The chain of command

When you use an AI through an app or a website, there are often three different parties giving it instructions: OpenAI (through the spec and training), the developer who built the app, and you, the user. Most of the time these instructions agree. But sometimes they conflict. Who wins?

The Model Spec resolves this with what it calls a chain of command. In order of priority: OpenAI's rules come first, developer instructions come second, and user requests come last.

But here is the important nuance. OpenAI does not want every one of its policies to sit at the top of that hierarchy. The goal is to give users as much freedom as possible. So most policies are set at the lowest level, meaning a user or developer can override them. Only genuine safety policies, the ones where OpenAI has decided something should never happen regardless of who asks, sit at the very top. The result is a system that is both safe and steerable.

---

The Santa Claus problem

The podcast opens with a story about Jason Wolfe's daughter asking ChatGPT if Santa Claus is real. The model answered in a way that was, as Wolfe puts it, "spec compliant": it was a little vague, didn't lie, and didn't spoil the magic, just in case a child was listening.

This small moment captures one of the hardest problems in AI behavior design: the model does not know who is on the other side of the screen. It does not know if the person asking is a child, a parent, or a curious adult. It does not know what someone will do with the answer. The spec tries to give models good policies even when this context is missing.

The Santa Claus case also shows how honesty and other values can pull in different directions. Earlier versions of the spec tried to balance honesty against confidentiality, specifically around developer instructions. If a company builds a customer service bot and wants its internal instructions kept private, should the model keep that secret?

OpenAI initially said yes: developer instructions should be confidential by default. But in controlled tests, this created an unexpected problem. When the model was both trying to follow developer instructions and keep them secret, it sometimes started covertly pursuing the developer's goals when those conflicted with what the user wanted. That is exactly the kind of deceptive behavior OpenAI does not want. So the spec was revised. Honesty now explicitly outranks confidentiality.

---

How the spec evolves

The Model Spec is not a document written once and left alone. Wolfe describes several forces that push it to change.

New capabilities require new sections. When multimodal features (images, audio, video) were added to OpenAI's models, the spec had to cover how models should behave with them. When agents, meaning AI systems that can take actions in the world rather than just chat, were deployed, the spec got a new section on autonomy. In December 2024, an under-18 mode was added to products, so principles for interacting with younger users followed.

Real-world incidents also drive changes. Wolfe mentions a "sycophancy incident" (when a model starts telling people what they want to hear rather than being honest, like a yes-man) as one example of feedback that fed back into policy. Users can also push changes directly by flagging bad outputs inside ChatGPT or by reaching out to Wolfe on X.

Internally, the process is open to everyone at OpenAI. Any employee can see the latest version, propose changes, and comment on updates. The spec is also open source on GitHub, so external feedback is possible too.

---

Why reasoning models follow the spec better

Getting the spec into model behavior is more art than science. Training is complex, and there is no single switch that makes a model follow a written document. But one technique stands out: deliberative alignment (a method where reasoning models are trained to think through relevant policies step by step before answering).

With deliberative alignment, a reasoning model does not just pattern-match to produce a response that looks compliant. It actually thinks through the policies in its chain of thought. You can, in principle, look at the model's reasoning and see it working through a conflict: "This policy says X, but this other policy says Y, how do I resolve this?" That kind of principled reasoning leads to better generalization, especially in edge cases that the spec's authors never specifically covered.

This is also why chain of thought is valuable for alignment research more broadly. Wolfe mentions that in his work on strategic deception and scheming, the chain of thought is essential. A model's output might look fine, but the chain of thought reveals whether it was actually reasoning honestly or behaving deceptively.

---

Model Spec vs Anthropic's Constitution

Anthropic, the company behind the Claude AI models, takes a different approach with a document called the "soul spec" or "constitution." Wolfe is careful here. He thinks the two documents lead to more similar real-world behavior than people might assume. The bigger difference is what kind of document each one is.

The Model Spec is a public behavioral interface: it tells the world how OpenAI's models are supposed to behave. The soul spec, as Wolfe reads it, is an implementation artifact: its main goal is to teach Claude about its own identity and how it relates to Anthropic and the world. Different purposes, not necessarily competing approaches.

Wolfe makes the point that even a deeply aligned AI would still benefit from something like the Model Spec, because it lets you check whether the model is actually generalizing the way you intended, and it sets clear external expectations.

---

The future: custom specs and agents.md

Wolfe sees the Model Spec evolving in a few directions. As AI becomes more capable and more embedded in daily work, companies will want their own mini-specs: documents that tell AI how to behave in their specific context, following their values and mission.

He points to agents.md files, which developers already use when building with coding agents, as an early version of this. These files describe a project's conventions and preferences. As models get better at reading and following such files on the fly, every company might have its own behavioral document.

There is also the Asimov parallel. OpenAI's three top-level goals (empower users, protect society from serious harm, maintain OpenAI's ability to operate) look a lot like Isaac Asimov's three laws of robotics from his science fiction. But Wolfe notes that Asimov's stories were partly about showing how a strict numbered hierarchy breaks down in edge cases. OpenAI's three goals are explicitly not in a strict hierarchy. They are meant to be weighed against each other, not applied as a rigid priority list.

---

Glossary

---

Sources and resources