(1) is usually true if you have good git history or commented code, without that you can just guess. LLMs do not reason, but they can be good at learning from patterns and replicating them in different contexts. So this still has some value and can save time.

Many programmers are nowadays using agents/copilots, and you can’t really distinguish if a code was “co-created” when reviewing a PR. All projects need to live with this fact nowadays, and it won’t go away but likely increase. Human review is still the best garbage filter we have.

For those worried about AI quality etc. Nothing gets merged without human reviews, proper commit structure, etc. So if what AI agent proposed works but it’s a crappy implementation it won’t get merged until it meets certain code quality standards.