LLM hallucinations in the wild: Large-scale evidence from non-existent citations

2026/05/08 by Zhenyue Zhao, Yihe Wang, Toby Stuart +3 · 37 voices · 1 citation
#cs.DL #cs.AI #cs.CY #physics.soc-ph

paper · pdf

Abstract

Large language models (LLMs) are known to generate plausible but false information across a wide range of contexts, yet the real-world magnitude and consequences of this hallucination problem remain poorly understood. Here we leverage a uniquely verifiable object - scientific citations - to audit 111 million references across 2.5 million papers in arXiv, bioRxiv, SSRN, and PubMed Central. We find a sharp rise in non-existent references following widespread LLM adoption, with a conservative estimate of 146,932 hallucinated citations in 2025 alone. These errors are diffusely embedded across many papers but especially pronounced in fields with rapid AI uptake, in manuscripts with linguistic signatures of AI-assisted writing, and among small and early-career author teams. At the same time, hallucinated references disproportionately assign credit to already prominent and male scholars, suggesting that LLM-generated errors may reinforce existing inequities in scientific recognition. Preprint moderation and journal publication processes capture only a fraction of these errors, suggesting that the spread of hallucinated content has outpaced existing safeguards. Together, these findings demonstrate that LLM hallucinations are infiltrating knowledge production at scale, threatening both the reliability and equity of future scientific discovery as human and AI systems draw on the existing literature.

Citations

• Hallucinated citations are polluting the scientific literature. What can be done?
• HalluCitation Matters: Revealing the Impact of Hallucinated References with 300 Hallucinated Papers in ACL Conferences
• Scientific production in the era of large language models
• Quantifying large language model usage in scientific papers
• AI models collapse when trained on recursively generated data

Cited by

• Detecting Hallucinated and Suspicious Citations: What Current Tools Can and Cannot Do

Discussions

• Two new facts stand out: 1. 85% of hallucinated citations in preprints are also in the subsequent journal version (thanks, peer review!) 2. Fake cites more likely to use the names of (male) scholars w [bsky, 629 points, 18 comments]
• “…hallucinated references disproportionately assign credit to already prominent and male scholars, suggesting that LLM-generated errors may reinforce existing inequities in scientific recognition…” 🧪 [bsky, 432 points, 13 comments]
• "We find a sharp rise in non-existent references following widespread LLM adoption, with a conservative estimate of 146,932 hallucinated citations in 2025 alone....hallucinated references disproportio [bsky, 76 points, 4 comments]
• 🚨 The growing KNOWLEDGE CRISIS.👇 LLM hallucinations in the wild: Large-scale evidence from non-existent citations arxiv.org/abs/2605.07723 [pre-print] "LLM hallucinations are infiltrating knowledge [bsky, 31 points, 2 comments]
• 'Hallucinated citations are more prevalent in work authored by researchers with little pre-2022 publication history. When fake citations occur, they disproportionately credit already established, high [bsky, 17 points, 0 comments]
• This is indeed a perfect opportunity for analyzing and automating citation checks with LLMs and scripts, IMHO. Kinda baffled that this is not already happening proactively. [bsky, 16 points, 2 comments]
• Quantifying the explosion of fake AI citations in the research literature--146,932 BS cites in papers from 2025 alone. Depressingly but unsurprisingly, social-science work had the most fake citations. [bsky, 14 points, 0 comments]
• "Together, these findings demonstrate that LLM hallucinations are infiltrating knowledge production at scale, threatening both the reliability and equity of future scientific discovery as human and AI [bsky, 13 points, 3 comments]
• "hallucinated references disproportionately assign credit to already prominent and male scholars, suggesting that LLM-generated errors may reinforce existing inequities in scientific recognition" arxi [bsky, 13 points, 1 comments]
• Hallucinated citations often prescribe authorship to already prominent and male scholars 🧪 arxiv.org/abs/2605.07723 [bsky, 10 points, 0 comments]
• Link to pdf file: arxiv.org/pdf/2605.07723 [bsky, 10 points, 1 comments]
• This preprint on hallucinated citations suggests a greater problem on SSRN (reaching 1.9% of total by August 2025) than arXiv, bioRxiv, or PMC (variously 0.2-0.4%). Intriguing if methods hold up; stud [bsky, 9 points, 1 comments]
• Tutkijat kävivät läpi 111 miljoonaa viitettä 2,5 miljoonasta tutkimusartikkelista (2020–2025). Suuren kielimallimullistuksen seurauksena tekaistut viittaukset lähtivät rajuun nousuun. Vuonna 2025 "hal [bsky, 7 points, 1 comments]
• Not great Bob arxiv.org/abs/2605.07723 [bsky, 6 points, 0 comments]
• the founder of arxiv just co-published a study finding ~147k hallucinated citations in 2025 papers, and they cluster around already-famous male researchers ... the machines are quietly building a fake [bsky, 6 points, 0 comments]
• @mblanchard.bsky.social : arxiv.org/abs/2605.07723 [bsky, 6 points, 1 comments]
• Hi Mel, is this related? This work came across my field of view very recently, and I already sent it to my "faculty list I'm bugging about not going further into University-sponsored LLM use for stude [bsky, 6 points, 1 comments]
• . . .these findings demonstrate that LLM hallucinations are infiltrating knowledge production at scale, threatening both the reliability and equity of future scientific discovery as human and AI syste [bsky, 6 points, 1 comments]
• LLM hallucinations in the wild: Large-scale evidence from non-existent citations arxiv.org/abs/2605.07723 [bsky, 5 points, 0 comments]
• LLM Hallucinations in the Wild [hn, 4 points, 1 comments]
• New Preprint: #LLM #Hallucinations in the Wild: Large-Scale Evidence From Non-Existent Citations (via #arXiv) arxiv.org/abs/2605.07723 #GenAI #LLMs #scholcomm #libraries [bsky, 3 points, 1 comments]
• New study of LLM hallucinations in citations using preprint articles from 2025: What discipline has the most fake citations? The social sciences, judging by SSRN, where nearly 2% of citations in 2025 [bsky, 3 points, 1 comments]
• From @retractionwatch.com & @arxiv-cs-cv.bsky.social | #LLM #hallucinations in the wild: Large-scale evidence from non-existent citations | #OpenScience #Ethics #Publications | 🧪 🔓| "conservative es [bsky, 3 points, 0 comments]
• Wenn während des Verfassens eines Papers ein LLM Zitat + Quelle ausdenkt, schafft es beides auch meist in die veröffentlichte Version. [2605.07723] LLM hallucinations in the wild: Large-scale evidence [bsky, 2 points, 0 comments]
• Happy Friday! arxiv.org/abs/2605.07723 [bsky, 1 points, 0 comments]
• Code ist ja Sprache und es kommt nicht auf Genauigkeit an, kann mir schon vorstellen, dass das hilfreich ist für jemanden der sich auskennt. Ich hab mir nur mal vbs für Excel-Shizzle machen lassen, fu [bsky, 1 points, 3 comments]
• First, they are NOT hallucinations. Machines and algorithms cannot hallucinate. They are faulty predictions degenerating information into falsehoods or false synthetic texts. (Doherty, forthcoming) [bsky, 1 points, 0 comments]
• “AI hallucinations in science are polluting academic research at massive scale, with 146,932 fake citations detected in 2025 alone.” [bsky, 0 points, 0 comments]
• https://bsky.app/profile/socrates17.bsky.social/post/3mlrsyea7kc2b [bsky, 0 points, 0 comments]
• LLM hallucinations in the wild: Large-scale evidence from non-existent citations. arxiv.org/pdf/2605.07723 [bsky, 0 points, 0 comments]
• "We find a sharp rise in non-existent references following widespread LLM adoption, with a conservative estimate of 146,932 hallucinated citations in 2025 alone." arxiv.org/abs/2605.07723 [bsky, 0 points, 0 comments]
• 147,000 hallucinated citations among 111m in 2.5m papers is actually much lower than I expected arxiv.org/abs/2605.07723 [bsky, 0 points, 0 comments]
• "Von den in den Preprints enthaltenen Halluzinationen finden sich 85,3 % auch in der veröffentlichten Version wieder".. Dies lasse vermuten, dass auch die aktuellen Prüfungsverfahren der Fachzeitschri [bsky, 0 points, 0 comments]
• paper arxiv.org/abs/2605.07723 [bsky, 0 points, 0 comments]
• arxiv.org/abs/2605.07723 [bsky, 0 points, 0 comments]
• Bookmarkin' for myself to read later. Hoo boy: arxiv.org/abs/2605.07723 [bsky, 0 points, 0 comments]
• https://arxiv.org/abs/2605.07723 [bsky, 0 points, 0 comments]

Related

• Hierarchical Memorization in Large Language Models: Evidence from Citation Generation
• Hallucinations in Bibliographic Recommendation: Citation Frequency as a Proxy for Training Data Redundancy
• Do Language Models Know When They're Hallucinating References?
• How unique are hallucinated citations offered by generative Artificial Intelligence models?
• HalluCitation Matters: Revealing the Impact of Hallucinated References with 300 Hallucinated Papers in ACL Conferences