Luckily, gemini catches a good amount of errors in PR reviews, less need for manual review unless you need to double check if the code structure and architecture is sane.
Depends on how good the human reviewers are. It's hard to give a thorough code review, you need to understand the code and requirements, pull the changes locally, manually test the PR, think about security, readability, catch line level bugs like bad conditionals, but also think about the overall structure (classes, patterns, systems). But this requires a lot of effort, especially with larger PRs and it's easy for things to slip through. Nothing is perfect, but you can think of AI review like a supercharged linter, it might not suggest an alternative approach, but it will point out some glaring omissions or unhandled exceptions, etc.
