What's interesting about this is that a human would hypothetically produce a similar error, but in practice would reject the question as beyond their means. I'd assume something about supervised learning makes the models overestimate their abilities. It probably learns that “good” responses attempt to answer the question rather than giving up.