146,900 Fake Citations Found in Scientific Papers as AI Use Spreads
Researchers auditing millions of scientific papers found an estimated 146,900 hallucinated citations across four major repositories in 2025 alone. The fake references appeared in papers on arXiv, bioRxiv, SSRN, and PubMed Central-spreading across many papers rather than concentrating in a few outliers.
Large language models generate text that sounds plausible but can be entirely fabricated. These systems predict the next word based on patterns in training data, not on actual facts. When researchers use AI to write papers without fact-checking the output, those hallucinations slip into the scientific record.
The Scale of the Problem
The audit examined 111 million references from 2.5 million papers. Researchers used automated checks and manual verification to identify citations that didn't match any real publication. Over 95% of references were successfully matched to actual sources.
The fake citations surged starting in mid-2024, after ChatGPT and similar tools became widely available. To isolate AI's role, researchers compared unmatched citation rates before 2023-when these tools didn't exist-against current rates. The difference was stark.
Early-career scientists and small research teams were most likely to include fake citations. Some of these researchers saw their output increase roughly threefold since AI tools became available.
Gaps in Quality Control
Preprint servers, journal editors, and peer review caught only a fraction of the errors. On arXiv alone, 78.8% of non-existent citations passed through moderation and appeared on the platform.
The hallucinated references showed a pattern: they disproportionately credited already prominent and male scholars. This suggests AI errors may reinforce existing inequalities in how scientific credit gets distributed.
Implications for Research
Scientific progress depends on building on verified prior work. Fake citations undermine that foundation. If researchers cite papers that don't exist, others may waste time trying to find them or, worse, build arguments on false premises.
The problem extends beyond academia. Hallucinations appear in government reports, legal filings, and news articles. Without intervention, the contamination of scientific literature could spread to policy decisions and public understanding.
Scientists need to verify AI-generated content before publication. Journals and preprint servers need stronger detection systems. The responsibility falls on researchers to check their tools' work-something many are not doing now.
To understand how AI systems generate these errors and best practices for using them responsibly in research, consider exploring Generative AI and LLM Courses or AI Research Courses.
Your membership also unlocks:
