Edge cases will usually be the ones to get through. Most developers don’t correctly write tests that exercise the limits of each input (or indeed have time to both unit test every function that way, and integration test to be sure the bigger stories are correctly working). Nothing about ai assist changes any of this.

(If anybody starts doing significant fully unsupervised “ai” coding they would likely pay the price in extreme instability so I’m assuming here that humans still basically read/skim PRs the same as they always have)