Hidden traps in submitted papers caught about 1% of AI-written peer reviews, but mostly the careless ones, so the true share is unknown.

This explores whether hidden watermarks planted in papers can tell us how many peer reviewers actually use LLMs, or only how many get caught. The short answer from the corpus is no. Watermarks give you a floor, not a measurement. ICML hid instructions inside submission PDFs. If a reviewer pasted the paper into an LLM, the model followed the hidden instruction and left a recognizable trace in the review. This flagged 795 reviews, about 1%, and led to 497 desk rejections. The chairs themselves say the method mostly catches careless users, and misses anyone who spotted and removed the instruction or rewrote the LLM's output How many peer reviewers secretly used LLMs despite the ban?. So the 1% counts reviewers who were careless. It does not count all the reviewers who used an LLM.

The more revealing evidence comes from a different method. ICML 2026 also ran a randomized experiment that assigned reviewers either a full ban or a limited-use policy. Substantial fractions of reviewers broke whichever rule they were given Does banning LLM use in peer review change review outcomes?. Set that next to the watermark's 1% and the gap is the point: the watermark catches a small, visible slice of the real behavior. The same experiment found that banning versus limiting LLM use barely changed scores or decisions. That suggests the prevalence number may matter less for review outcomes than the policy debate assumes.

Other kinds of detection don't fill the gap. Readers with ML expertise couldn't reliably tell LLM-written abstracts from human ones, and tended to assume a human was involved Can readers tell LLM abstracts from human ones?. ICLR 2026 treated statistical LLM detectors as too unreliable to act on alone. It routed their flags to area chairs as one input among several, and saved hard enforcement for something checkable: fabricated references How can conferences detect and handle LLM misuse in peer review?. The pattern is to enforce only where the evidence can be confirmed, and not to treat any detector as a census.

The twist is that the hidden-prompt technique cuts both ways. The watermarks work because LLMs obey instructions buried in a document. Authors have used the same weakness: 18 arXiv manuscripts contained concealed prompts telling AI reviewers to praise the paper, a practice now classed as questionable research conduct Are hidden AI prompts in preprints a deceptive research practice?. LLM judges are also easily swayed by fake authority signals and polished formatting Can LLM judges be fooled by fake credentials and formatting?. So a measurement tool built on prompt injection depends on the same flaw that makes LLM reviewing open to manipulation. It also teaches careful reviewers what to look for, which means it is likely to undercount more over time.

If you want a truer picture of prevalence, the corpus points toward large-scale statistical and outcome-based analysis rather than catching individuals. One study of 125,000+ reviews looked at LLM-assisted reviewing across the whole review pool. It found that the apparent favoritism of LLM-assisted reviewers toward LLM-written papers disappears once you control for paper quality Do LLM reviewers actually favor LLM-written papers?. The lesson there is about method more than about cheating: even population-level signals can mislead unless you account for confounders. The corpus doesn't contain a validated true-prevalence estimate for LLM use in peer review, and that absence is itself telling.