I don't believe AI can perform these tasks but I would be happy to be proven wrong.

I am testing a similar idea in the next weeks. So far I have an adversary reviewer, that works better than a simple reviewer, and complains the hell out of the code. 😂

Because LLM's don't "use" software like humans do. LLM's aren't trying to move thumbs around on a touchscreen screen, or push a mouse to click things. LLM's don't get annoyed when they have to wait ~30 seconds for something to load. LLM's don't have good aesthetic sense to tell when a font selection doesn't work, or a UI is too crowded / buggy. LLM's can't tell when an icon is too small to be seen.

I'm not a LLMs fan at all, but you are wrong in the majority of what you are saying. LLMs have a great visual analysis capabilities, can play with applications and have huge knowledge about performances, usability and accessibility indicators. They only have some limits in simulating inaccurate or multiple touch interactions.