From 2020 to 2025, scholarly articles totaling 2.5 million were examined by analysts at Cornell, UCLA, and Berkeley. These documents contributed a citation count of 111 million. Data originated in prominent archives - arXiv, bioRxiv, SSRN, and PubMed Central being among them. Attention shifted toward references that lacked confirmation in standard indexing systems. Tools like Semantic Scholar, OpenAlex, and Google Scholar failed to validate certain paper titles. Scrutiny centered on these unverifiable instances. Work unfolded without reliance on assumed accuracy.
Instead, gaps in traceability became the point of departure.
Midway through 2024, a noticeable spike emerged in made-up citations. This shift came alongside broader adoption of advanced language software - systems initially built for drafting text but now able to produce full reference lists. Although such tools speed up writing tasks, they sometimes invent scholarly sources that sound real yet lead nowhere.
A paper called "LLM Hallucinations in the Wild" traced this pattern directly to how these models operate when asked to cite materials. Because false references mimic genuine ones so closely, spotting them becomes difficult without careful checking.
Surprisingly, the investigation reveals fabricated citations appear beyond clearly dishonest work. These false references turn up across credible-looking documents, implying certain authors include AI-suggested sources without checking them first. What stands out is how casually unverified material slips into accepted formats.
Most current safety measures faced questions about how well they work. The research showed that close to 78.8% of made-up citations got through arXiv’s review process without detection. Even after some bioRxiv papers appeared in journals listed by PubMed Central, around 85.3% still kept their false references unchanged.
A study appearing in The Lancet highlighted recurring issues in biomedical literature.
Over 4,000 false references turned up in nearly three thousand reviewed articles from 2023 through early 2026. Papers drawn from that span showed a sharp climb in made-up sources. While just one in 2,828 works contained such problems at the start, the proportion jumped - by early 2026, it was one out of every 277. Growth like this signals deeper cracks forming beneath the surface.
One concern gaining traction: false references might cycle back into AI training data once they land in shared digital archives. Because these inaccuracies can persist, journals are being pushed toward using software checks on citations prior to accepting articles.
As artificial intelligence plays a larger role in research tasks, closer scrutiny seems less like an option and more like a necessity. Some now see automated validation not as extra effort but as basic hygiene in scholarly communication.
