Because AI is really good at generating code that looks good on its own, on both first and second glance. It's only when you notice the cumulative effects of layers if such PRs that the cracks really show.

Humans are pretty terrible at reliable high quality choice review. The only thing worse is all the other things we've tried.

Because they can produce magnitude more code than you can review. And personally I don't want to review _any_ submitted AI code if I don't have a guarantee that the person who prompted it has reviewed it before.

It's just disrespectful. Why would anyone want to review the output of an LLM without any more context? If you really want to help, submit the prompt, the llm thinking tokens along with the final code. There are only nefarious reasons not to.