You have to trust OpenAI whether they cryptographically signed the response or not, so they could just tell you what model was used to generate the response.
That's true! OpenAI would/could lie. Even so, a signature would solve a lot of problems: aggregators/resellers rewriting the {model: X, ...} in the payload, staged benchmarks (there's currently no way to prove a dataset's answers actually came from the model claimed), MITM... Of course, it's not that dire, and there's no incentive for OpenAI to look at it. But I was kinda shocked to find out there's no way to guarantee something this basic (though maybe it's not as simple as I think). Who knows? It could add a lot of latency. I'd rather think it's more a lack of will