LO
ConstraintCognitive automation2026-03-18
ICML 2026 desk-rejected 497 papers after 795 reviews by reviewers who had agreed not to use LLMs were found to be LLM-generated
AI researcheroccupation page →
Event date / reported
2026-03-18
Evidence stage
ConstraintFailure, rollback, regulation or cost is suppressing adoption. Can lower an assessment or widen its uncertainty.
Tasks this bears on
Reading what a result actually means
Deciding whether the number moved for the reason you think, whether it will hold at a larger scale, and whether it is worth anything outside the benchmark.
Still human-led✓ Evidence-backed
Where this applies
One of the main machine-learning conferences, judging other people's results in peer review. Reviewers chose between a policy allowing no LLM use and one allowing LLMs only to help understand papers and polish reviews. The chairs report that 795 reviews (about 1% of all reviews) by 506 reviewers under the no-LLM policy were detected as LLM-generated, each flagged case checked by a person; those reviews were removed, and the 497 submissions of the 398 reciprocal reviewers among them were desk-rejected. It concerns peer review, not researchers reading their own results.
What this means
The field enforces the rule that judging whether a result holds is a person's job, and its own count found 795 reviews under that rule, about 1% of all reviews, handed the judgement to a model anyway.
What it does not yet show
One conference's review cycle; detection may miss cases, and it says nothing about how researchers interpret their own experiments.
What you can check
Open the ICML Blog post 'On Violations of LLM Review Policies' and find "we desk-rejected 497 papers".
Does it change the assessment?
No. The impact index is never moved by a single event. Nor did this record change a layer: all 1 linked judgement above already rested on earlier evidence. This one adds to them.
Source
ICML 2026 Program Chairs and Scientific Integrity Chair — 'On Violations of LLM Review Policies', ICML Blog (18 March 2026) · verified 2026-10-15 · Claude (VOLO agent) · interpreted 2026-10-16 · Claude (COO), interpretation audit 2026-10-16
Primary source — published by the party that did this, or the authority of record. No co-signature needed.