Who pays the review cost? 239 reviewers say passing CI doesn't make an agent PR reviewable
A survey of 239 practitioners finds reviewers don't reject AI-authored PRs outright. They gate them on scope, rationale, validation beyond CI, responsiveness and a named owner, and call the missing work reviewability debt.
Source: Who Pays the Review Cost? Triage, Fairness, and Accountability in AI-authored Pull Requests
Md Shamimur Rahman, Khairul Alam, Banani Roy and Chanchal K. Roy (University of Saskatchewan) surveyed 239 practitioners from 31 countries who review code, most of them heavy reviewers of AI-authored PRs (78.2% had reviewed at least 10, 49.4% more than 40). Since AI PRs became common, 77.8% report more PRs needing review, and 74.1% say review communication has changed moderately or significantly.
The sharpest number is about trust in green builds. Asked whether they would accept a CI-passing change without further explanation, reviewers were less willing for AI-authored PRs on all eight change types. Even low-risk cleanup dropped from 86.2% to 73.6%. The biggest drop was for refactors that claim no behavior change: 52.3% for human authors, 19.7% for AI authors. Before reviewing, 66.5% want CI or tests passing first and 56.9% want a clear statement of intent, and 41.8% want evidence a known human will maintain the code. The most common reason to defer was a vague, generic AI-written description (64.9%), followed by very large undivided PRs and missing tests for behavior changes (both 51.0%).
Reviewers did not want blanket bans. For suspected drive-by AI PRs from newcomers, only 12.6% would close unmodified AI output; 55.6% would first ask the contributor to show they understand the change. The credibility signals that mattered were behaviors in the current review: professional response to feedback (82.4% rated highly influential) and testing evidence beyond "it works for me" (81.2%), far above an active GitHub profile (23.4%). The authors name the gap reviewability debt: the explanation, validation and ownership work a reviewer must absorb when a PR looks finished but nobody human stands behind it.
Why it matters for code review
This is the human side of running agents around the clock. Generating more PRs overnight is cheap; reviewer attention is not, and reviewers say plainly that they shift from mentoring into "audit mode" when a submission offloads its validation onto them. Respondents asked for rules over vigilance: evidence bundles matched to the change type, caps on open AI PRs per contributor, routing to domain owners, and feeding review corrections back into agent instructions. A harness that ships modest, owned, well-evidenced PRs earns review; one that ships volume runs up debt. Read beside the enterprise 2× mandate study, where AI PRs spent 20% longer between first review and merge, and Nachuma on actionable review loops.