AI-authored pull-request review: human participation can mean agent steering

Original: Duma et al., These Aren’t the Reviews You’re Looking For, May 4, 2026, with a public replication package. Observational GitHub analysis of AIDev-tagged author accounts, review comments and selected repositories; manually filtered mislabelled bots. Among 33,596 AI-authored PRs in repositories with at least 100 stars, 61.38% had no recorded review activity, 22.6% agent-only review, 15.9% some recorded human participation. These fractions say nothing about silent maintainer inspection.

The within-same-repository subset matters more for comparison: 9,616 agent-authored and 5,574 human-authored PRs. Both had almost identical rates of observable any human participation, 30.1% and 30.8%, respectively. But among reviewed PRs, agent-authored changes had human-only review 8.08% versus 25.21% for human-authored changes; mixing human and agent reviewers was more common on agent changes. About one-quarter of the human comments on agent PRs were classified as commands steering agents, versus about 1.6% on human PRs. Review-comment categories were regex-validated on 800 sampled comments but the residual ‘human review’ class includes ‘LGTM’ and workflow messages, not necessarily technical examination. The body text accidentally states 40.34% mixed review in the larger subset, while its table gives 31.09% (4,034/12,975); use counts/table, not the stray prose figure.

Selection is not random and neither matched issue difficulty nor reviewer minutes, failures, defect detection or later outcome is observed. Do not turn the 84% of all PRs lacking a recorded human review in the broad set into an industry rate or assume the same for merged work. Takeaway for a coordinator: distinguish human test/decision, human steering and agent-only review, with an inspectable disposition of each accepted warning. See clock comparison and warning-gate trace.