I think part of it is that most users at some point encounter an error message that is just straight up wrong. For example, a login page that says "wrong password" when in reality the user is typing EXACTLY what they typed on account creation, but the site silently truncated the password. Even one such frustrating experience is enough to teach many users that as soon as they see any error message, they should stop trusting anything the system tells them, including the error message. It's extremely difficult to rebuild user trust after this sort of UX contract violation, particularly because less technical users don't mentally differentiate separate computer systems. All the systems are just "the computer."

Also arguably the users are kind of right. An error indicates that a program has violated its invariants, which may lead to undefined behavior. Any output from a program after entering the realm of undefined behavior SHOULD be mistrusted, including error messages.

Personally I think it's kind of the opposite. AI is like having a personal assistant that aced the first two years of undergrad courses in every subject under the sun, but isn't capable of original reasoning on the frontier of knowledge. AI rewards users who have a good high level map of different areas of knowledge and the verbal/analytical skills to ask good questions. It used to be that to approach an engineering problem in an unfamiliar technical domain, it might take weeks or months of background study to even know what search term to use. I've found that AI reduces that knowledge bootstrap to hours or days. I've been able to work in problem domains I never would have touched otherwise.