What is AI code review?

AI code review is the use of a language model to review code changes before they ship. Code review is the check in which a colleague reads a proposed change and decides whether to approve it; the change is usually a pull request, a request to merge someone's work into the shared codebase. With AI, the model reads the same change and leaves comments: it flags likely errors, proposes fixes and summarizes what changed. It can comment on every pull request on its own or assist a human reviewer. The issue became urgent with coding agents: when machines write more code, the bottleneck moves to review, which is also where an organization decides what it accepts and who answers for it. An AI reviewer helps triage the work, but it cannot take responsibility for an approval, and every comment it leaves takes time from whoever has to read and resolve it.

See also: AI code smell · Recurring defects in LLM and agent written code: it passes the demo, but maintenance costs more than expected.

What the studies show

Cihan and colleagues (ICSE 2025) studied a company that uses a reviewer built on a language model, analyzing 4,335 pull requests, 1,568 of which the tool also reviewed (Cihan et al., 2025). Of the automated comments, 73.8% were resolved, but average pull request closure time rose from 5 hours 52 minutes to 8 hours 20 minutes overall, though not in every project, and developers reported faulty reviews, unnecessary corrections and irrelevant comments. The AIDev dataset contains over 456,000 pull requests opened by agents; the agents work faster than people, but their pull requests are accepted less often (Li, Zhang and Hassan, 2025). Among 567 pull requests written with Claude Code, 45.1% of the merged ones were revised before merging, a share similar to that of the same authors' human-written pull requests (Watanabe et al., 2025). Google Cloud's DORA research program, which works from survey data, points the same way: higher AI adoption goes with more delivery throughput but also more instability, and time saved by authors often gets spent by reviewers (DORA, 2026). An AI reviewer can flag the surface features of AI code smells, but it brings two risks: the person who trusts its silence (automation bias) and the model that goes along with a confidently written change (sycophancy).

An enterprise example

A logistics company adopts a coding agent and, at the same time, an AI reviewer that comments on every pull request. Open pull requests pile up right away, and its two senior developers spend their days resolving automated comments, many of them irrelevant. A habit soon forms: if the AI reviewer finds nothing, the change gets approved quickly. That is how an error in the shipping-rate calculation reaches production: the code was clean, but it broke a business rule the model had no way of knowing. The fix changes where the tool steps in, without removing it: the AI reviewer runs before the pull request opens and the author must resolve its comments, generated changes are split into small batches, and critical modules require approval from a named owner.

Why it matters for decision makers

Review is where an organization declares that code is acceptable, and someone has to sign off on that declaration. GitHub's documentation says Copilot reviews are comments by default and do not count toward required approvals, but a feature in public preview lets administrators make them count the same as a teammate's. Turning that on is a governance decision before it is a setting. Judge the tool by review time, reworked changes and incidents, not by resolved comments. And where approval becomes a rubber stamp, nobody really understands the code being merged, and the team builds up cognitive debt.