A review of 22 paper submissions across three computer vision and machine learning conferences — NeurIPS, WACV, and TerraBytes (an ECCV geospatial workshop) — found that 15 submissions (68%) contained fabricated citations, fabricated author lists for existing papers, and/or were clearly written by large language models (LLMs), according to a report published July 30, 2026 (GeospatialML).

The reviewers, who examined submissions in mid-2026, checked 11 papers for fabricated citations and found 6 (55%) contained them. Separately, of 11 submissions checked for fabricated authors or LLM-generated writing, 9 (82%) were flagged. The problem spanned all three venues. At WACV, 3 of 4 reviewed submissions had fabricated citations and 3 of 4 had fabricated authors or LLM-generated writing. At NeurIPS's Position Paper track, 2 of 2 reviewed submissions had fabricated authors or LLM-generated writing. At NeurIPS's Datasets and Benchmarks track, 2 of 5 reviewed submissions had fabricated citations or LLM-generated writing. At TerraBytes, 1 of 2 reviewed submissions had fabricated citations and 4 of 5 had fabricated authors or LLM-generated writing.

The sample size is small and the selection method is not described as random, so these rates should be read as a signal, not an estimate of overall prevalence. What makes the signal hard to dismiss is corroboration from larger-scale studies across multiple venues and publishers.

A Nature analysis found that at least tens of thousands of 2025 publications probably contain invalid AI-generated references (Nature). Zhao et al. estimated roughly 146,900 hallucinated citations — references that look real but point to papers that do not exist — appeared across arXiv, bioRxiv, SSRN, and PubMed Central in 2025 alone, and found that 85.3% of hallucinated references in bioRxiv preprints persisted into their published versions, indicating that peer review is not catching 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 accepted proceedings at top-tier venues. Ansari (2026) analyzed 100 hallucinated citations from papers accepted at NeurIPS 2025 and found that every one made it past three to five expert reviewers. In total, 53 papers containing hallucinated citations were accepted at NeurIPS 2025, roughly 1% of acceptances, and remain in the proceedings (Ansari, arXiv).

LLM-generated content is not confined to submissions. Pangram found that 21% of ICLR 2026 reviews (15,899 reviews total) were fully AI-generated, and over half had some form of AI involvement (Pangram). ICML 2026 took an active countermeasure, embedding prompt-injection stings — hidden instructions planted inside submission text that cause an LLM reviewing the paper to reveal itself by responding to the instruction — to detect LLM use in reviewing. The approach identified 795 reviews (approximately 1% of all reviews) by 506 unique reviewers that used LLMs despite an explicit policy prohibiting it (ICML Blog).

The author of the July 30 review report released a bib-audit skill on GitHub for detecting fabricated citations (GitHub), making the methodology used in the audit available for other reviewers.

The 85.3% persistence rate from Zhao et al. deserves particular attention. If peer review is failing to catch hallucinated citations at that rate, the burden of detection is shifting downstream to readers, meta-analysts, and anyone who treats a published reference list as evidence that a cited source exists. The Lancet data quantifies the trajectory: a roughly sixfold increase in the incidence rate of papers with fabricated references over two years. The NeurIPS 2025 data from Ansari adds that even multi-reviewer evaluation at selective venues does not reliably catch fabricated citations, since all 100 sampled hallucinated references survived review.

ICML 2026's prompt-injection approach is the most concrete operational response to date, and it produced a measurable detection rate. But it targeted reviewers, not submitters. The July 30 audit data, alongside the Nature and Lancet findings, indicates that the submission side of the pipeline remains under-addressed. A reviewer manually checking 22 submissions and finding fabricated citations in over half of those examined is a rate that, even with a small and non-random sample, is consistent with the larger-scale findings.

The bib-audit skill released alongside the report is a practical tool, but it places the detection burden on individual reviewers who are already performing unpaid labor. Conference-scale infrastructure for automated citation verification does not yet appear to exist at any of the venues named. Whether the ICML prompt-injection model can be adapted to submissions, or whether venues will adopt systematic citation validation, is an open question.

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