This is one of the most honest takes on AI-generated PRs I've read. The "AI can produce plausible but wrong code at scale" point is underappreciated — the danger isn't that AI code is obviously broken, it's that it looks clean so reviewers lower their guard. You've described exactly what's happening on teams right now.

The dedup logic point is critical and something most people building these tools get wrong. A reviewer that posts the same comment 3 times trains engineers to ignore everything it says. The credibility model matters as much as the accuracy model.

Your layered defense framing (author intent → AI first pass → human final decision) is how teams should be thinking about this. The problem we keep seeing is that as generation speed increases, teams skip layer 2 entirely and go straight to rubber-stamping at layer 3.

Curious how your team handled the signal-to-noise calibration over time — specifically the severity thresholds. Getting engineers to trust "high" actually means high seems to be the hardest adoption problem in practice.