Because he is smart enough to use the existing (frontier) tools to get good results and create a sort of collaborative environment that is novel for research maths.
Though by your "Perelman and Wiles didn't need "AI" assistance" comment, you'd surely be there on the sidelines to ridicule them for each and every single one. I guess maybe that's where your concerns are coming from?
I can practically see how these concerns of yours would suddenly evaporate if they started using self-hosted models instead... ... yeah, right, who are we kidding?
People who don't take advantage of the best available tools and techniques don't get to that level to begin with.