Rather than directly admit fault they'll regularly respond in subtle (moreso o3) to not so subtle roundabout ways that deflect blame rather than admit direct fault, even when it's an inarguable factual error about even conceptually non-heated things like API methods.
It's an annoying behavior of their models and in complete contrast to say Anthropic's Claude which ime will immediately and directly admit to things it had responded incorrectly about when the user mentions it (perhaps too eagerly).
I have wondered if this is something its learned based on training from places like Reddit, or if OpenAI deliberately taught it or instructed via system prompts to seem more infallible or if models like Claude were made to deliberately reduce that aspect.
I don't know whats better here. ChatGPT did have a tendency to reply with things like "Oh, I'm sorry, you are right that x is wrong because of y. Instead of x, you should do x"
Human-level AI is closer than I'd realised... at this rate it'll have a seat in the senate by 2030.