Log every wrong AI review comment, dismiss it with a short reason, and review the log monthly to tune the tool or narrow its scope. Ignoring false positives quietly teaches the team to ignore all of them, including the real ones.
Expect mistakes
Vendors say so themselves. GitHub's documentation for Copilot code review states that it will make mistakes and is not guaranteed to spot all problems, and advises validating its feedback carefully. Plan your process around that, not around hoping it will not happen.
The three kinds of wrong
- False positive. It flags code that is fine.
- Incorrect claim. It describes behavior that is not true, such as an API that does not work as stated.
- Harmful suggestion. A fix that introduces a bug or weakens security.
The third is the one to fear. Reviewers should always read a suggested change, not click apply on reflex.
A lightweight loop
- Dismiss the comment with a one line reason, so the author and later readers see why.
- React or label it as wrong, using a convention your team picks.
- Add the case to a shared log: file type, comment, why wrong.
- Each month, group the log by cause.
What the log tells you
Patterns usually show up. One rule may fire on generated files. Another may misread a framework you use. A third may fail on a language feature. Each pattern suggests a fix: exclude a path, add repository instructions if your tool supports them, or turn off a category.
Do not train people to ignore
If a bot posts twenty comments and nineteen are noise, the twentieth gets skipped. Cut volume. Move mechanical rules into linters. Narrow the bot to what it does well. Google's practice of labeling minor points as nits applies here too, since authors need a quick way to tell what matters.
Do not hide misses either
False negatives matter more than false positives. When a human finds a bug the bot missed, log that too. Over time, the ratio of useful comments to wrong ones and the list of misses give you evidence for keeping, limiting or dropping the tool.
Escalate to the vendor when it is systematic
If a wrong claim repeats, report it through the vendor channel with a reproducible example, and note the date and tool version in your log.
Share the findings
Post the monthly summary where the team will see it. When people see that wrong comments lead to tuning, they report them more. When reports vanish into a void, they stop.
Quick check
Should wrong bot comments just be ignored?
No. Dismiss them with a reason and log them, so the team can tune the tool and keep trust in the real findings.
Which kind of wrong comment is most dangerous?
A harmful suggestion, such as a fix that adds a bug, which is why every suggested change needs a read before applying.
Sources
Last reviewed: 19 October 2026.
