Researchers have built an AI code reviewer that checks whether an AI coding agent's output actually matches what the developer asked for, rather than just flagging style nits.
The system, called ARCTIC, does three things: it infers the intent behind a change from conversation logs and metadata, measures how far the agent's actual code drifted from that intent using a backtranslation technique, and ranks which parts of a diff most need a human's eyes. The team built it after analyzing 18,000 code reviews and distilling six recurring review themes. In offline tests, intent prediction scored 0.86 F1, drift detection nearly matched human raters' judgments (a 0.907 agreement score), and the spotlight feature beat a baseline AI reviewer at rating code quality by 2.4x while using a fifth of the tokens. In a live rollout, the drift scores cut code misalignment by another 5.76 points, engineers approved 90.2% of the intent predictions, and the team reports zero defects traced back to diffs that were self-reviewed with the tool since launch.
That gap - between what AI reviewers currently flag and what human reviewers actually care about - is the real story here. Most AI code review products lean on the same instincts as a linter: naming conventions, formatting, best practices. Correctness, security, and performance, the things that actually cause outages, get comparatively little attention. As coding agents ship more code than any team can manually review, tools that triage for what matters rather than what's easy to detect become the bottleneck worth solving.
Worth remembering: these are the paper's own numbers, measured against one baseline reviewer and one internal rollout. Nobody outside that rollout has kicked the tires yet.