Can AI-generated papers pass peer review undetected?

Sakana AI reports that one of the three papers its AI Scientist-v2 generated passed double-blind peer review at an ICLR 2025 workshop. The papers were "entirely generated end-to-end by AI, without any modifications from humans": the system proposed the hypothesis, designed and ran the experiments, analyzed the data, and wrote every word of the manuscript. The human team chose only the broad topic and which three papers to submit. The accepted paper, "Compositional Regularization: Unexpected Obstacles in Enhancing Neural Network Generalization", reports a negative result and averaged 6.33 from reviewers. Sakana says this ranks "approximately 45% of all submissions" and sits above the workshop's average acceptance threshold. Only one of the three was accepted.

The evidence comes from a designed experiment rather than an ordinary submission. Reviewers were told that 3 of 43 papers might be AI-generated, but not which ones. Under a protocol agreed in advance, any accepted AI paper would be withdrawn and desk-rejected, and kept off OpenReview's public forum, because the AI and scientific communities have not decided whether AI-generated manuscripts belong in the same venues. The organizers also skipped a meta-review because they already knew of the experiment, so the 6.33 is a set of reviewer scores with no final decision behind it. The stated reason for withdrawal is about norms, not about the quality of the reviews.

Read against the library, this is the verification gap in concrete form. The note on Can AI verify research outputs as fast as it generates them? argues that generation runs ahead of proof. Here the artifacts cleared a human review, and the builders' own reading then found a citation error (an LSTM network credited to Goodfellow (2016) rather than Hochreiter and Schmidhuber (1997)) and judged that none of the three met the bar for an ICLR main-track paper. That fits the worry in Does polished writing actually signal better quality work?, though the excerpt shows only scores and does not show that reviewers were swayed by polish. It also contrasts with Can inference scaling help reviewers catch errors humans miss?: that note describes a machine check finding flaws that passed human review, while this excerpt does not say whether the reviewers saw the errors the authors later found. The ICML experiment in Does banning LLM use in peer review change review outcomes? measured reviewer behavior under LLM rules. This excerpt measures how reviewers score papers when AI authorship is possible but unidentified.

The excerpt does not establish much beyond its own account. It reports one workshop, three submissions and one acceptance, with no reviewer comments, no independent replication, and no analysis of how reviewers reached their scores. The acceptance-rate comparison it offers (20-30% at main conferences, 60-70% at workshops) is Sakana's own framing, and the builders call the work preliminary. The implication, at the strength the evidence allows, is narrow: an end-to-end AI pipeline can produce a manuscript that clears one workshop's double-blind review at a score its builders describe as above threshold. It does not show that the work would clear a main-track bar, which Sakana's own review says it did not, and it does not show that review scores track merit. The forecast that such systems will reach top journals is Sakana's prediction, not a finding.