Accuracy measures how often AI-generated review findings or code suggestions are correct against the team’s review criteria. It matters because automation can reduce review effort while still requiring human judgment. section 13.1.2 section 13.4
Treat accuracy as a team-managed confidence level rather than a reason to automate blindly:
- Low confidence: use AI for initial analysis, inspect what it flags and misses, avoid blocking pull requests, and rate its output when possible. section 13.4
- Medium confidence: take suggestions more seriously, but allow human reviewers to override incorrect findings. section 13.4
- Set thresholds for acceptable incorrect and correct suggestions, then record expectations in the team working agreement. section 13.4
A concrete accuracy warning
One reviewer-assisting model translated natural-language review comments into code modifications with 31% accuracy, showing why generated changes still need review. section 13.1.2
