Calculations are fairly easy to fix, though AI may still make order of operations errors -just as humans. The difference is that AI will note the differences and offer two or more "drafts" /versions of the answer.
The bigger problem is when people expect AI to output "truth." For one thing, because people who create AIs are just as subject to confirmation bias as anyone else. Iow, they may disapprove of what their AI outputs for many reasons.
My eldest daughter works on a Bard team. One problem she notes is that some responses have to be translated from one cultural idiom to another. For ex., she had to ask a Japanese team leader to explain why "warm belly" had a specific relation to Japanese seasons. I don't recall it exactly. But, just imagine that much of the data on the internet is slang or idiomatic. So, the AI has to have different ways to say the same thing to different people. As in, you ask a question, then ask the AI to explain the answer another way.
Of course, that also means the AI can "understand" things differently as well. That might make it more prone to hallucinations, too.
