The analysis of 111 million references from 2.5 million papers found unmatched or non-existent citations rising after widespread adoption of large language models.

Citations across arXiv, bioRxiv, SSRN, and PubMed Central (2020–2025) were checked against Semantic Scholar, OpenAlex, and Google Scholar, revealing many that could not be verified.

Fabricated citations risk compromising clinical guidelines and systematic reviews; publishers are urged to implement automated reference verification before acceptance.

Current safeguards fail: about 78.8% of fake citations bypass arXiv moderation, and 85.3% of hallucinated references in bioRxiv preprints appear in final PubMed Central-indexed journals.

A Cornell-UCLA study identified 146,900 AI-generated fake citations across four major databases, highlighting reliability concerns for cited sources in scientific work.

The overall takeaway is that AI-produced fake citations are infiltrating biomedical and broader scientific literature, underscoring the need for stronger verification and safeguards.

A large-scale study from Cornell, UCLA, and UC Berkeley estimated around 150,000 fabricated references entering scientific records in 2025, largely moving from preprints to peer-reviewed journals.

Surge in AI-generated hallucinations post-2022 is linked to the evolution of AI tools from writing aids to citation-generation engines around mid-2024.

AI-generated fake citations are not confined to a few papers but are spread across many, signaling widespread reliance on AI-generated references without proper verification.

A Lancet audit of 2.5 million biomedical papers found a sharp rise in fabricated references, with over 4,000 fake citations across 2,810 papers observed from 2023 to early 2026.

The deterioration is quantified: about 1 in 2,828 papers in 2023 contained at least one fabricated citation, rising to 1 in 458 in 2025 and 1 in 277 by early 2026.

ArXiv announced steps to ban authors who submit work with hallucinated citations or unvetted AI content, aiming to curb the problem.