Reviewers flag fake authors in ML papers — venues accept them as orals anyway

Two machine-learning reviewers describe a peer-review season drowning in AI-generated submissions. Across 22 papers they were assigned for NeurIPS, WACV, and a geospatial ECCV workshop, 15 (68%) contained fabricated citations, invented author lists grafted onto real papers, or writing that was unmistakably LLM-produced — including a 53-page NeurIPS submission stuffed with hallucinated jargon and stray notes left in by the model. In two cases the reviewers recognized real papers whose true authors had been swapped for plausible-sounding fakes; they recommended rejection and alerted organizers, only to see both accepted for oral presentation on the condition that the bogus references simply be fixed.

The anecdotes sit atop a growing body of measurement. A Nature analysis estimated tens of thousands of 2025 papers likely carry invalid AI-generated references, and an audit by Zhao et al. put the count near 146,900 hallucinated citations for the year, concentrated among early-career researchers and small teams. Crucially, review is not catching them: 85.3% of hallucinated references in bioRxiv preprints survived into the published versions, and a Lancet audit found the share of biomedical papers with at least one fabricated reference rising six-fold in two years, reaching one in 277 by early 2026.

The rot runs both directions. Pangram found 21% of ICLR 2026 reviews were fully AI-generated, ICML 2026’s prompt-injection stings caught reviewers who had pledged not to use LLMs, and Li et al. showed adversarially reworded abstracts can inflate AI-reviewer scores by more than a point without altering the underlying science. The authors’ blunt takeaway: papers whose own citations are fabricated signal that the writers never read their own work, and should be desk-rejected rather than waved through — a stance that cuts against venues currently treating hallucinated references as a cosmetic fix.