Review fatigue

When AI code review becomes noise

Review bots that comment on everything train people to ignore them. Why AI code review gets noisy, and how to get signal: fewer claims, shown on the code, that you decide on.

A familiar story: an architectural pull request, seventy-two comments, one human in the thread. Nobody reads them all. After a few weeks, nobody reads most of them.

Why it gets noisy

Review bots flag what could be wrong, not what is. Style nits, vague "consider adding tests", rewrites that do not fix a bug. Each one costs a moment of attention, and attention is the thing a review is short of.

The result is alert fatigue. The real problem, when there is one, arrives looking exactly like the forty comments before it.

Signal is a short argument

What a reviewer needs is the opposite of a pile of comments: a few claims about the change, ordered by importance, each with the evidence beside it, and a clear question at the end.

That is a deck. With deck, the agent that made the change shows you three or four groups, not seventy findings. Each one is a claim with its lines lit. You read the argument in a few minutes, and you disagree where it matters.

Practical ways to cut the noise

• Turn off style comments in bots; let a formatter handle style.
• Ask for decisions first: what was chosen, and what was the alternative.
• Review one concern at a time.
• Keep comments for things a human must decide.