It can answer any question but the answer won't necessarily be right.
Lazy people will think it is a simple matter of adding some little piece to it that can classify its right and wrong answers but that is a problem like trying to get a bubble out from under a rug.
The book
https://en.wikipedia.org/wiki/G%C3%B6del,_Escher,_Bach
has a set of dialogues about some people struggling furiously to build an impossible machine who take a long time to understand what they're trying to do is impossible. I've talked to people who see that book as a rebuttal to the old symbolic A.I. which was fundamentally based on logic, but the constraints that it talks about apply to any technology because they are about the very concept of logical truth, which is something that you have have to master to correctly answer a wide range of questions. Neural networks do not escape the results of Turing, Godel, Tarski and such, unless they don't promise to get the right answers.
Back when I was a Chuunibyou and really into things like Bigfoot and E.S.P. I remember reading something by Edgar Cayce about how there were two pathways to the answers you get doing channeling or pendulum dowsing (which I liked to do back then) There was one for the true answers and one for the false and you could become a good psychic if you could tell the difference. Of course you can't. Cayce was wrong about many things like much of the U.S. sinking beneath the waves, the U.S. government reinventing the death ray that sank Atlantis, etc.