I predict the way AI will be useful in science from the perspective of mathematics is by figuring out combinatorially complex solutions to problems that would otherwise not be interesting to (or far too complex to be comprehended by) humans. With such capabilities it could be imagined then that the AI will be useful for designing super materials, or doing fancy things with biology / medicine, and generally finding useful patterns in complex systems.
However, they at most would be the heuristic function of a search mechanism. A good heuristic, but heuristic at most. For search we need to identify when to abandon a path and which other entry point is promising. I'm not sure our current techniques are good for this kind of problems.
I think it will be even more helpful to know a simple proof doesn't exist because AI has tried for long enough and didn't find it. Once people know there is no easy proof of say Collatz or Twin Primes Conjencture those will not be as alluring to waste your time on.
But abstract mathematics doesn’t care about solutions to problems; it cares about understanding problem spaces. I do not think current AI helps with that.
Problems like the one discussed also aren’t interesting to applied mathematicians, either, because of lack of applications.
But yes, if this kind of AI produces new materials, solves diseases, etc. they will be very useful. We wouldn’t care whether they arrived at those solutions through valid reasoning, though. A sloppy AI that has better ‘guesses/intuitions’ than humans or that can guess and check ‘guesses/intuitions' for correctness faster would be very useful.