AI is only good at a subset of problems. However in that subset it is often very good. There is a lot of "low hanging fruit" in that subset and we can learn a lot quickly by having AI work in that area.
We do of course need to watch for hallucinations, but I in general expect AI to be very good at theoretical physics. AI can take a lot of known equations and prove/propose (these are different things!) generalizations to that may or may not match reality. AI can suggest experiments to see if the predictions match the real world. Maybe string theory can finely make a non-trivial prediction that we can test in the real world...
However AI is terrible for other parts of physics (at least so far) and those are also important we shouldn't lose track of them despite the excitement that we can get from progress in things AI is good at.