1 hour ago
3
ARTICLE AD BOX
Suggests Several Researchers Copying Them Without Check
BENGALURU: A surge of AI-generated fake citations has entered technological literature, with a caller survey estimating that astir 1.5 lakh fabricated references slipped into the grounds successful 2025, astir moving from preprints into peer-reviewed journals.This is the cardinal uncovering of a large-scale survey by researchers astatine Cornell University, UCLA and UC Berkeley, which analysed 111 cardinal citations crossed 2.5 cardinal probe papers published betwixt 2020 and 2025 connected arXiv, bioRxiv, SSRN and PubMed Central. The study, titled 'LLM hallucinations successful the wild', tracked citations whose titles could not beryllium verified against large world databases, including Semantic Scholar, OpenAlex and Google Scholar.
By comparing post-2022 trends against pre-ChatGPT mistake baselines, researchers isolated the apt publication of AI-generated hallucinations to the surge. The steepest emergence began astir mid-2024, astir 18 months aft ChatGPT's nationalist release, arsenic AI tools evolved from penning assistants into citation-generation engines. The contamination is not concentrated successful evidently fraudulent papers. Researchers recovered that fake references are typically sprinkled sparsely crossed different morganatic manuscripts, suggesting galore researchers are copying AI-generated citations without verifying them.
Existing safeguards are failing. Nearly 78.8% of fake citations passed arXiv moderation, and among bioRxiv preprints aboriginal published successful PubMed Central-indexed journals, 85.3% of hallucinated references made it into the last published versions. Researchers pass the occupation whitethorn present beryllium self-reinforcing: arsenic fabricated references embed themselves successful open-access repositories and citation databases, aboriginal AI models trained connected that corpus hazard absorbing and reproducing the aforesaid hallucinations.Study successful Lancet excessively warns In a abstracted survey titled "Fabricated citations: an audit crossed 2.5 cardinal biomedical papers" published successful The Lancet, researchers recovered a crisp emergence successful fabricated citations successful biomedical probe papers. The study, conducted by researchers from Columbia University and different institutions, analysed biomedical papers published betwixt 2023 and aboriginal 2026. It recovered much than 4,000 fabricated references embedded crossed 2,810 peer-reviewed papers. The audit recovered the complaint of fabricated references roseate dramatically implicit the three-year period. In 2023, astir 1 successful 2,828 papers contained astatine slightest 1 fabricated citation. By 2025, the fig had worsened to 1 successful 458 papers, and by aboriginal 2026, it had climbed to 1 successful 277 papers. One of the astir striking examples cited successful the survey progressive a 2025 insubstantial successful an open-access oncology diary connected ureteroileal surgical techniques.
Researchers recovered that 18 of the paper's 30 verified references (60%) were fabricated. The authors linked the surge partially to wide adoption of LLMs which are known to "hallucinate" fake citations. Warning that fabricated citations could compromise objective guidelines and systematic reviews, researchers urged publishers to present automated notation verification systems earlier papers are accepted for publication. Nearly 98% of the affected papers had not faced immoderate steadfast enactment astatine the clip of the audit, the survey noted.
