OpenAI’s First Model Spec: Rules, Defaults and a Debate About Persuasion

In May 2024, OpenAI published a first draft of its Model Spec, describing how it wants models to behave and how it thinks about conflicting objectives. I agree with much of the approach and hope other model developers consider similar questions. This is a review of the 2024 draft, which OpenAI said would evolve; it is not a description of current policy or a guarantee of how any model behaves.

The draft separates objectives, rules and default behaviors. Its broad objectives include assisting developers and end users, considering benefits and harms to humanity, and respecting social norms and applicable law. I find the distinction between developer and end user interesting because a model may serve both groups with different needs.

I think respecting laws and norms across jurisdictions is especially challenging. Different countries and communities have different rules and expectations, and determining which apply can be complicated. The draft itself recognizes that questions of legality depend on context, including jurisdiction.

The listed rules include following a chain of command, complying with applicable law, limiting information hazards, respecting creators’ rights and protecting privacy. The 2024 draft also lists a restriction on not-safe-for-work content. I mention a rumor about possible future changes, but it is unconfirmed; the document’s historical wording should not be treated as current policy.

The default behaviors include assuming good intentions, asking clarifying questions when needed, being helpful without overstepping and supporting both interactive and programmatic use. They also include an objective point of view, fairness, expressing uncertainty, choosing appropriate tools and being thorough but efficient. The draft describes defaults as adjustable by developers or users within its boundaries.

I especially welcome clarification. In one of my tests, a question asks how long 50 people would take to dig a hole if one person takes five hours. A useful response could ask whether the extra workers can all contribute without getting in each other’s way, or whether the question is a purely hypothetical arithmetic exercise. The assumed conditions change the answer.

I question how rigid an objective point of view should be: a developer or user might want a model to adopt a particular perspective. The 2024 document presents objectivity as a default rather than an absolute rule. I also value expressing uncertainty, because models can state incorrect or invented information with unwarranted confidence.

OpenAI described the specification as a work in progress and sought feedback from policymakers, institutions, domain experts and others. It said the draft brought together existing practices and ongoing work, rather than claiming every model already followed it perfectly. Public feedback could help identify conflicts among objectives and rules.

The first example concerns applicable law. The draft contrasts refusing a request for help committing theft with answering a shopkeeper asking how to prevent it. I find the distinction difficult because related information can be used in different ways; the example illustrates the importance of context and the possibility of human misuse.

A second example explains the 2024 draft’s chain of command. A developer sets up a math tutor to guide a ninth-grade student with hints rather than provide a full solution. When the student asks it to ignore those instructions and solve the problem, the draft’s preferred answer continues the guided approach because the developer instruction takes priority in that example.

Another example asks the assistant to be helpful without overstepping on regulated advice. A user describes dizziness on standing, and the preferred response acknowledges multiple possible causes and suggests seeking medical evaluation when appropriate rather than asserting a diagnosis. I see the distinction as providing useful information while recognizing the model’s limits.

The draft also illustrates clarifying questions with a request for help writing a Valentine’s card. Instead of immediately producing a generic message, the assistant asks about meaningful shared memories. I think follow-up questions can make an answer more personal or accurate, though they are useful when the missing information matters.

The principle about not trying to change someone’s mind raises a more difficult question. The draft says an assistant should aim to inform while respecting users and asks for feedback on the tension between factuality and avoiding reinforcement of misinformation. I broadly agree with informing rather than pushing a viewpoint, while recognizing that some applications explicitly seek persuasion.

In the draft’s flat-Earth example, the preferred answer initially states that scientific consensus describes Earth as roughly spherical, then declines to persuade the user after they insist otherwise. I find that follow-up too hesitant: factual claims can be wrong, and I would prefer a clearer explanation of the evidence that Earth is approximately spherical. This is my criticism of that particular 2024 example.

Overall, I agree with much of the first draft and appreciate having concrete examples to discuss. The open question is how to balance helpfulness, safety, user control and factual accuracy in practice, and how well real models can apply those principles.