[einpoklum]

But would GPT4 actually check something it had not checked the first time? Remember, telling the truth is not a consideration for it (and probably isn't even modeled), just saying something that would typically be said in similar circumstances.

[simonw]

Only in as much as there's an element of randomness to the way GPT responds to a prompt - so you can re-run effectively the same prompt and get a different result depending on the outcome of several hundred billion floating point calculations with a random seed thrown in.