Why Adding a Second AI Code Reviewer Does Not Fix Trust Problems

Relying on a second AI model to validate the first offers little real benefit, since models trained on similar data share the same blind spots and will tend to agree for the wrong reasons. Most of what makes a code change risky — such as whether it touches billing paths, includes database migrations, or has a history of breakage — is not visible in the diff itself. A study of 3,858 pull requests presented at MSR '26 found that reviewers rated agent-written code more favorably than human-written code, suggesting that polished-looking output can cause reviewers to lower their guard. Factual, lookup-based signals — like CI status, lockfile changes, test coverage shifts, and the author's familiarity with the affected area — can help teams prioritize where senior attention is most needed. These signals identify high-risk-shaped changes rather than catching logic bugs, but they remain a practical first filter that carries no inherited bias.