要点
- Five values: faithfulness to human intent, harm minimization, honesty and transparency, facilitating human agency, and care; the Spec describes target behavior the models do not yet fully reflect.
- Adds or strengthens an updated chain of command, careful use of context, defaults on human agency and social norms, and side-effect communication; retires most detailed style guidance.
- Points to new training methods for learning these values and why principles matter, with results deferred — this post itself carries almost no new behavioral metrics.
On 29 October 2026, OpenAI updated its public Model Spec, folding a more itemized behavior guide into five underlying values. As models get smarter and take more consequential actions, literal instruction-following is not enough: they need judgment where instructions are silent, while staying faithful to human intent and oversight. This is a governance-document reorg, not a model release.
[1]Concrete textual changes include rewriting faithfulness in an assistance-games spirit with an updated chain of command and consider context carefully; new defaults on supporting human agency and following relevant social norms, plus expanded side-effect control and communication; retiring most fine-grained style guidance while keeping stronger defaults such as approaching questions objectively; and referring to the model as “GPT” rather than “the assistant.” OpenAI also says it is exploring training methods so models learn why principles matter and how to apply them in new situations — but the eval results are not in this post.
[1]What is checkable is the document: the value list, the section map, and the self-limiting claim that the Spec is aspirational. That is more auditable than vague “we care about safety” language. What is weak is equally clear: no before/after numbers on value-following, instruction-conflict resolution, or user preference after style rules were retired. “Future updates” leave the causal chain in the promise zone. Readers can verify what OpenAI says; they cannot verify from this post that models already behave better under the new framing.
[1]Steelman: writing values and trade-offs explicitly, then iterating training, is more criticizable — and improvable — than black-box RLHF slogans; retiring style rules in favor of user preference also fits “do not silently aestheticize for the user.” But if conflicts among the five values still resolve mainly through internal chains and defaults, while outsiders only read prose, the accountability interface stays narrow. Putting “care” and “agency” into defaults can also create new paternalistic failure modes before they are well measured — the post’s own warning against quietly steering users concedes that risk.
[1]For developers, chain-of-command and context defaults shape whether product-level instructions can override platform boundaries. For enterprise buyers, a Spec bump is compliance narrative; contracts that cite Spec versions without binding behavioral evals purchase documentation, not guarantees. For researchers of manipulation and dependency, the agency and care clauses supply new attack surfaces and eval outlines.
[1]Over the next six months, watch whether value-learning training yields reproducible conflict-scenario evals; whether developers can actually override style-like defaults; and whether the public feedback channel rewrites specific clauses. If the Spec grows into a moral-philosophy outline while behavior curves stay flat, this reorg’s main output is clearer accountability text — not a more aligned model.
[1]