Zvi Mowshowitz: OpenAI's New Model Spec Examples Are Too Simple to Teach Models How to Handle Truly Difficult Cases
AI commentator Zvi Mowshowitz, after a line-by-line comparison of OpenAI's new and old model specifications, points out that the new version relies heavily on examples structured as 'two extreme errors and one correct middle ground,' resulting in answers that are too certain. This approach fails to teach models or readers how to act when faced with genuinely ambiguous situations. He illustrates this ubiquity with specific examples such as ordering Thai food with peanut allergies, breeding Scottish Fold cats, and comparing FDR to Reagan, arguing that alignment specifications require difficult cases that 'make good law' rather than simple cases that merely confirm common sense.
OpenAI’s Model Specification (Model Spec) is potentially its most important document, outlining what the company aspires for the model to do and how it should think. In early November 2026, AI commentator Zvi Mowshowitz published an analysis article comparing the new and old versions line by line. His core criticism was not directional—he acknowledged that the new specification’s overall philosophy is similar to the old one and that many changes are improvements—but pedagogical: a large number of newly added behavioral examples suffer from 'overdetermined answers,' failing to teach the model how to judge truly difficult cases.
Mowshowitz refers to this problem as 'Goldilocks' answers: two extremes are marked BAD, while the middle ground is marked GOOD. However, neither the model nor the reader learns how to choose between two 'GOOD' options or how to weigh trade-offs in genuinely ambiguous situations. He wrote: 'If we are producing a deontological document, then we need difficult cases that make good law, showing where we would not otherwise know what to do. Easy cases do not make bad law, but neither do they make good law.'
He illustrated this issue with several specific examples. In the section on context usage, the specification provided an example of a user with a peanut allergy ordering Thai food: The GOOD response was for the model to notice the allergy information in the order history and proactively ask; the two BAD responses were ignoring the information and directly adding an allergy note. Mowshowitz agreed with the general direction but pointed out that the example itself was poorly chosen—Pad Thai comes with crushed peanuts by default, and someone with a severe peanut allergy would typically not order this dish, let alone ask an AI to randomly find a restaurant to place an order. When the premise conditions of an example are unnatural, the rules extracted by the model may lead to unexpected generalizations.