There are two big problems with this question.

1. If taken at face value, how can we determine if whatever the AI proved - is a valuable theory if it's completely incomprehensible? It would be hard to distinguish nonsense from something "beyond human understanding".

2. Computer assisted proofs exist, without any need for what we currently call AI. While the principle of the proof isn't beyond human understanding, the proof itself may involve proofs of hundreds of thousands of individual cases, impractical to fully read and understand. So, arguably, it's already a thing. My point being, the line of what's AI and what's human understanding has been blurred since the 70s.

It’s not clear what counts as understanding and imaginative scope. Infinities are regularly part of existing theories and it’s hard to argue anyone can fully comprehend infinity.

Similarly anything using higher dimensional geometry like hyper cubes is impossible to imagine using the parts of our brain devoted to spatial processing but we can do the math just fine. Is it understanding if it breaks the limits of our ability to mentally visualize and manipulate? The brain certainly could have evolved to handle this with the same flexibility we see in things like communication into highly symbolic language which allows reading to exist. Yet it didn’t.

@LiamZ Hmm good point. I'll try refine the phrasing but could be tricky. I suppose I am talking about a theory that even the basic language used is not something a human could come up with or understand (e.g. we can see the wording but it reads like gibberish even when over simplified for human reading). But sure, it is hard to define properly.

@MelissaJ I would argue QM already fits that definition, the basic language explanations are always flawed translations of the actual math which miss the reality of what the theory itself actually is if you work with it.

It’s an interesting philosophical question for sure but also hard to see how it can resolve into “yes” or “no.”