A reviewer checked 22 research papers submitted to three computer science conferences — NeurIPS, WACV, and TerraBytes — and found that 15 of them (68%) contained made-up citations, fake author names, or text that was clearly written by an AI tool like ChatGPT, according to a report published July 30, 2026 (GeospatialML).

The reviewers examined the submissions in mid-2026. Of 11 papers they checked for fabricated citations, 6 (55%) had them. Of 11 papers checked for fake authors or AI-generated writing, 9 (82%) were flagged. The problem showed up at every conference they looked at. At WACV, 3 of 4 reviewed submissions had fabricated citations and 3 of 4 had fabricated authors or AI-generated writing. At NeurIPS's Position Paper track, both reviewed submissions had fabricated authors or AI-generated writing. At NeurIPS's Datasets and Benchmarks track, 2 of 5 reviewed submissions had fabricated citations or AI-generated writing. At TerraBytes, 1 of 2 reviewed submissions had fabricated citations and 4 of 5 had fabricated authors or AI-generated writing.

The sample size is small and the selection method was not random, so these numbers are a warning sign rather than a precise measure of how common the problem is. What makes the warning hard to ignore is that larger studies from multiple sources are finding the same pattern.

A Nature analysis found that at least tens of thousands of publications from 2025 probably contain invalid AI-generated references (Nature). Zhao et al. estimated that roughly 146,900 hallucinated citations — references that look real but point to papers that do not actually exist — appeared across four major research repositories in 2025 alone. They also found that 85.3% of these fake references in biology preprints survived all the way into the final published version, meaning the peer review process did not catch them (Zhao et al., arXiv). An audit in The Lancet covering 2.5 million biomedical papers found the share of papers with at least one fabricated reference rose from 1 in 2,828 in 2023 to 1 in 458 in 2025 (The Lancet).

The problem has also reached papers that were accepted at top-tier conferences. Ansari (2026) analyzed 100 hallucinated citations from papers accepted at NeurIPS 2025 and found that every single one made it past three to five expert reviewers. In total, 53 papers containing hallucinated citations were accepted at NeurIPS 2025, roughly 1% of all acceptances, and they remain in the conference proceedings (Ansari, arXiv).

AI-generated content is not limited to submitted papers. A company called Pangram found that 21% of reviews written for ICLR 2026 (out of 15,899 reviews total) were fully AI-generated, and over half had some form of AI involvement (Pangram). ICML 2026 fought back by planting hidden instructions inside submission text. If a reviewer was using an AI to write their review, the AI would follow the hidden instruction and reveal itself. This approach caught 795 reviews (about 1% of all reviews) by 506 different reviewers who used AI despite a policy explicitly prohibiting it (ICML Blog).

The author of the July 30 review report also released a free tool on GitHub that checks for fabricated citations (GitHub), making the method used in the audit available to other reviewers.

The 85.3% persistence rate from Zhao et al. deserves particular attention. Peer review is the system where other experts check a paper before it gets published. If that system is failing to catch fake citations at that rate, the job of spotting them falls to readers, researchers who compile studies, and anyone who assumes that a citation in a published paper means the source actually exists. The Lancet data shows the trajectory: a roughly sixfold increase in papers with fabricated references over just two years. The NeurIPS 2025 data from Ansari adds that even multiple expert reviewers at highly selective conferences do not reliably catch fabricated citations, since all 100 sampled fake references survived review.

ICML 2026's hidden-instruction approach is the most concrete step any conference has taken so far, and it produced a measurable result. But it caught reviewers using AI, not people submitting AI-written papers. The July 30 audit data, alongside the Nature and Lancet findings, suggests that the submission side of the problem remains under-addressed. A reviewer manually checking 22 submissions and finding fabricated citations in over half of them is a rate that, even with a small and non-random sample, lines up with what the larger studies found.

The citation-checking tool released alongside the report is useful, but it puts the burden on individual reviewers who are already working unpaid. None of the conferences named appear to have large-scale automated systems for verifying citations. Whether the ICML approach can be adapted to catch problems in submissions, or whether conferences will adopt systematic citation checking, is an open question.

What is not in dispute is that fabricated citations and AI-generated text are now showing up at measurable rates across multiple levels of academic publishing, from small workshop submissions to main conference proceedings, and that current review processes are not reliably catching them.