Are AI code reviews getting worse?

GitHub argues that when using AI to review pull requests, explain diffs, or diagnose issues, accuracy and conciseness matter more than long, polished responses.

Overview

• Bigger AI models aren’t guaranteed to be better at code review.
• Recent benchmarks suggest some newer frontier models may produce longer, highly formatted reviews while still missing fundamental bugs.
• For AI-assisted PR review and debugging, verbosity (for example, “six fancy headings” in a review) can be a warning sign to double-check correctness.
• The focus should be on:
  • Conciseness (clear, minimal output)
  • Accuracy (catching real issues rather than sounding confident)
  • Signal over formatting (useful findings vs. presentation)