A new study by researchers from Cornell and UCLA has uncovered a growing problem in scientific publishing: AI-generated fake citations. Analyzing over 111 million references from 2.5 million papers, the team identified 146,900 citations that do not correspond to any real publication—hallucinations produced by large language models like ChatGPT and Gemini. These false references were found across four major scientific repositories: arXiv, bioRxiv, SSRN, and PubMed Central, which host early versions of research papers before formal peer review.

The study highlights a sharp increase in non-existent citations following the widespread adoption of AI tools in 2023. While some unmatched references were due to typos, many were entirely fabricated, suggesting researchers are using AI to draft citations without proper verification. This trend threatens the integrity of the scholarly record, which relies on accurate referencing for peer review and cumulative knowledge.

In response, arXiv has announced stricter policies, including banning authors who submit papers with hallucinated citations. Experts warn that unchecked AI-generated content adds noise to the scientific corpus, making it harder to identify valid research and potentially misdirecting future studies. The findings underscore the need for vigilance as AI becomes more embedded in academic workflows.