AI Code Review Is Still a Review

Automated AI tools for code review can effectively handle syntax, style consistency, vulnerability patterns, surface-level logic errors, and test coverage signals, but they fall short in verifying code intent, business logic, and compliance with specifications. The reliance on AI for both generating and reviewing code can create a "circular trust problem," where AI models trained on similar data converge on incorrect solutions that appear correct due to consensus, leading to increased production incidents. To address this, human judgment remains crucial for ensuring that code meets business requirements and specifications, which automated reviews cannot reliably check. Implementing a human-improved review spec as an anchor can provide a necessary sanity check, and Aviator's upcoming Aviator Verify aims to bridge the gap by parsing code and running deterministic checks against acceptance criteria to ensure alignment with approved specifications.