Hallucinated citations in scientific papers have surged dramatically since the rise of AI tools. At least 1,46,932 fabricated references generated by artificial intelligence entered the scientific record in 2025 alone, with the vast majority of those detected in preprints surviving peer review and making their way into journal articles, reported TOI.
That is the key finding of a large-scale study conducted by researchers from Cornell University, University of California, Los Angeles and University of California, Berkeley, who analysed 111 million citations across 2.5 million research papers published between 2020 and 2025 on arXiv, bioRxiv, SSRN and PubMed Central.
The study, titled “LLM hallucinations in the wild”, tracked citations whose titles could not be verified against major academic databases such as Semantic Scholar, OpenAlex and Google Scholar. By comparing post-2022 trends with pre-ChatGPT error baselines, the researchers isolated the likely role of AI-generated hallucinations in the sharp increase, reported TOI.
The findings were striking. By August 2025, hallucinated citation rates had climbed to nearly 2% in SSRN papers, 0.4% in arXiv, 0.3% in PubMed Central and 0.2% in bioRxiv, with monthly fake citation estimates touching 8,140 in PubMed Central alone.
The steepest increase began around mid-2024, roughly 18 months after the public release of ChatGPT, as AI tools evolved from writing assistants into citation-generation engines, reported TOI.
Researchers noted that the contamination was not limited to obviously fraudulent papers. Instead, fake references were often sparsely distributed across otherwise legitimate manuscripts, suggesting that many researchers may be copying AI-generated citations without properly verifying them.
The issue appeared to disproportionately affect certain groups. Authors linked to hallucinated citations were generally less experienced, but their publication output grew rapidly — increasing 3.13 times faster on SSRN and more than doubling on bioRxiv compared with matched peers by 2025.
Solo researchers and smaller teams were also overrepresented. When hallucinated references pointed to real scientists, they tended to favour prominent scholars — those cited had 68.8% more prior publications and 58.3% more citations than average.
Existing safeguards appeared inadequate. Nearly 78.8% of fake citations passed arXiv moderation, and among bioRxiv preprints later published in PubMed Central-indexed journals, 85.3% of hallucinated references remained in the final published versions.
The researchers warned that the problem could become self-reinforcing. As fabricated references become embedded in open-access repositories and citation databases, future AI models trained on those datasets may absorb — and reproduce — the same hallucinations.
Study in The Lancet also raises concerns
In a separate study titled “Fabricated citations: an audit across 2·5 million biomedical papers” published in The Lancet, researchers reported a sharp increase in fabricated citations in biomedical research papers.
The study, conducted by researchers from Columbia University and other institutions, analysed biomedical papers published between 2023 and early 2026. It identified more than 4,000 fabricated references embedded across 2,810 peer-reviewed papers.
The audit found that the rate of fabricated references rose sharply over the three-year period. In 2023, roughly one in 2,828 papers contained at least one fabricated citation. By 2025, the figure had worsened to one in 458 papers, and by early 2026, it had climbed further to one in 277 papers.
One of the most notable examples highlighted in the study involved a 2025 paper in an open-access oncology journal on ureteroileal surgical techniques. Researchers found that 18 of the paper’s 30 verified references — or 60% — were fabricated.
The authors linked the surge partly to the widespread adoption of large language models (LLMs), which are known to “hallucinate” fake citations.
Warning that fabricated citations could compromise clinical guidelines and systematic reviews, the researchers urged publishers to introduce automated reference verification systems before papers are accepted for publication. The study also noted that nearly 98% of the affected papers had not faced any publisher action at the time of the audit.
