There’s a simple point to be made: if no one in the world understands an AI proof, what’s the point?
Maybe another AI could use the proof.
How would we know it was correct?
If you feed an AI nonsense in its training data, it will generate nonsense