Or at least this was true until recently. GPT-5 is consistently delivering more coherent and better working UIs, provided I use it with shadcn or alternative component libraries.
So while you can generate a lot of code very fast, testing UX and UI is still manual work - at least for me.
I am pretty sure, AI should not run the show. It is a sophisticated tool, but it is not a show runner - not yet.