Do AI code reviews actually catch real bugs?

Been running an AI reviewer on our PRs for a month. It flagged a real race condition we'd missed, which was impressive, but it also left 40 comments about naming and "consider extracting this function." Half the team now skips its output entirely. I'm torn, because the signal is there if you dig, but the noise makes digging expensive. Curious how others tune this: restrict it to security or logic checks