You finish a piece of work. You ask AI to review it. It says "looks good." You publish.
It wasn't good. You had no way to know.
That's the problem with adversarial review. You tell the AI to "act as a security reviewer" or "play the skeptic." It gives you feedback. Some is useful. Most of it you can't trust — not because the AI is malicious, but because you have no idea why it said what it said.
"Your experiment design looks solid." Based on what?
"Consider adding error handling." Which errors? Why here? Is this real analysis, or did the model just pattern-match "code review → suggest error handling"?
