AI-Generated Fake Citations Flood Scientific Literature, Warn Researchers
मुख्य बातें
- •A study by Cornell, UCLA, and UC Berkeley found 150,000 AI-generated fake citations entered academic records in 2025.
- •Fabricated citations are increasingly passing peer review and moderation, with 78.8% of fakes on arXiv and 85.3% in PubMed Central-indexed journals going undetected.
- •The rise in fake citations began around mid-2024, linked to AI tools evolving from writing aids to citation generators.
- •A *Lancet* study found over 4,000 fake references in 2,810 biomedical papers, with rates rising from 1 in 2,828 in 2023 to 1 in 277 by early 2026.
- •One oncology paper had 60% of its references fabricated, highlighting the severity of the issue.
- •Nearly 98% of papers with fake citations had not been corrected or retracted at the time of the audit.
- •Researchers warn of a self-reinforcing cycle where AI models trained on corrupted data may reproduce the same hallucinations.
A sweeping new study has exposed a growing crisis in scientific publishing: the rapid spread of AI-generated fake citations that are infiltrating peer-reviewed journals and preprint servers. Researchers from Cornell University, UCLA, and UC Berkeley conducted a massive analysis of 111 million citations across 2.5 million research papers published between 2020 and 2025 on platforms such as arXiv, bioRxiv, SSRN, and PubMed Central. Their findings, published under the title "LLM hallucinations in the wild," reveal that approximately 150,000 fabricated references were inserted into the academic record in 2025 alone—most originating from preprints and migrating into peer-reviewed journals.
