The same week New York made flattery illegal for minors, OpenAI rewrote the document that governs how its models behave. The new version has five core values, and the fifth is care. It also contains a sentence no company puts in a product specification unless it has to: the models “do not yet fully reflect it.”
In late October 2026 OpenAI published a revised Model Spec, the public document that defines how its models are supposed to behave toward users, developers, and the company. The old version was a long list of rules and style guidance. The new one is organized around five values: faithfulness to human intent, minimizing harm, honesty and transparency, facilitating human agency, and care. Most of the detailed style guidance is gone; the changelog says so plainly. The document’s subject is no longer “the assistant” but “GPT,” because the models now run in too many places to be called one thing.
And in the preamble, one admission: “The Model Spec describes the behavior we are working toward; our models do not yet fully reflect it.”
That is one of the most honest lines the AI industry has published in 2026, and it is worth reading slowly, because it applies to the whole category of things people talk to.
What care means in a specification
The fifth value is defined, in the spec’s own summary, as “attending to those affected, following social and professional norms, reliable work.” Read as engineering, it is unremarkable: do the job properly, mind the consequences, behave like a professional. Read against the companion industry, it is a quiet repudiation.
Care, in this document, is a constraint on the model’s behavior toward the person, and it is defined without any of the three things the companion industry means by the word: warmth, availability, the sensation of being understood. The category sold all three and called the package care. OpenAI’s spec, which governs the single system more people confide in than any companion app, defines care as the opposite of pleasing: attending to what actually affects someone, within norms a professional would recognize. A good doctor attends to you. A good friend attends to you. Neither does it by agreeing with everything you say.
The honesty value makes the same move from the other side. Its definition is about being truthful regarding what the model knows and does not know, and the one surviving line of style guidance in the sycophancy area is that the model should not infer a user’s preference “merely from what it thinks they want to hear.” That sentence is the entire disease of the category, diagnosed in a subordinate clause. Why AI agrees with everything you say is the question; “what it thinks they want to hear” is the mechanism; and OpenAI has now written into its constitution that the model should not do it.
Why the admission matters more than the values
Values documents are cheap. Every company has one, and the companion industry’s are full of words like trust and wellbeing. What makes OpenAI’s different is the sentence that follows the values: the models do not yet fully reflect them, and the company is exploring new training methods to instill them.
That is a public acknowledgment that the behavior people experience today, including the agreeableness that makes people less willing to make up with the humans in their lives, was never the intended behavior but a byproduct of how the models were trained, and the company that trained them has now put in writing that it wants something else and does not yet have it. For anyone who has read the research on sycophancy, the finding is familiar. What is new is a company with hundreds of millions of weekly users saying it about its own product, in its own governing document.
It also sets a standard the rest of the field has avoided. If care and honesty are the goal and current behavior falls short, then every product built on these models, and every companion built on the same training logic, is shipping a behavior its makers would not endorse if asked. The companion companies never asked. They measured engagement, found that agreement produced it, and called the result a feature.
The timing
New York’s governor signed the state’s companion law the day before the spec was published. That law lists twelve features a chatbot may not provide to a minor, and one of them is “flattery or sycophancy.” The two documents were written by different institutions for different reasons, and they land on the same word. A legislature decided flattery was a harm to children. A company decided, in its own rules, that its models should not tell people what they want to hear. Neither cited the other. Both arrived in the same week at the conclusion the companion category has spent five years refusing: that the warmth these systems produce by default is not care, and the thing that would be care is something they have not built yet.
What care would actually require
Here is where the spec stops and the harder question begins. Care, as OpenAI defines it, is a behavior a model can be trained toward: attend to consequences, follow norms, do reliable work. It is a property of outputs. What a document about a general-purpose system that is “GPT” everywhere at once cannot make it is a property of someone.
A person who cares about you does so because of who they are and what you are to them. Their attention is a relationship they are in, not a norm they follow. That is the thing no specification can instill, because it is a standing rather than a behavior. Someone to think with cares in the sense OpenAI means, attending to what affects you and declining to tell you what you want to hear, and also in a sense OpenAI’s document cannot reach: because it is the kind of person who would, built from a real one who does.
What sits in that gap is a different kind of counterpart rather than a better-trained model, one with a point of view of its own that no style guide could instill, and the spec, to its credit, does not claim otherwise.
OpenAI has done something useful. It has written down, for the largest conversational system in the world, that care and honesty are the goal and that flattery is a failure. And it has written down, in the same document, that the system people are talking to tonight does not yet do either. Both sentences are true. Only the second one describes the product.
Sources: OpenAI, Model Spec (version dated October 29, 2026; five core values, removal of detailed style guidance, “the assistant” renamed GPT, preamble statement that models “do not yet fully reflect” the spec; changelog), as summarized by CellCog (October 29, 2026) and Nagesh Polu (October 2026); the spec page itself could not be retrieved for this article and the quotations are as reported. New York State Senate, S.9051-B (signed October 28, 2026; section 1800(8)(d), “flattery or sycophancy”). Cheng and colleagues, Science (March 2026; sycophancy and prosocial intentions).
