A new study titled “LLM hallucinations in the wild: Large-scale evidence from non-existent citations” has raised alarm over the increased adoption of Large Language Models (LLMs) in research papers.
According to the study LLMs generate information which seems to be true but are actually false yet the usage and and consequences of this hallucination problem remain poorly understood in the real world.
The researchers from Cornell University, University of California Los Angeles, Tsinghua University and University of California Berkeley analyzed 111 million references across 2.5 million papers in arXiv, bioRxiv, SSRN, and PubMed Central and they found a sharp rise in non-existent references following widespread LLM adoption, with an estimated 146,932 hallucinated citations recorded in 2025 alone.
“These errors are diffusely embedded across many papers but specially pronounced in fields with rapid AI uptake, in manuscripts with linguistic signatures of AI-assisted writing, and among small and early-career author teams,” the study revealed.
The study mentioned that these hallucinated references disproportionately assign credit to already well-known and male scholars, suggesting that these AI errors may reinforce existing inequities in scientific recognition. It further highlighted that LLM hallucinations are increasingly infiltrating knowledge production, which could make future scientific discovery unreliable and unequal.
“The propensity of LLMs to generate plausible but false information, often referred to as hallucination, remains an unresolved challenge even in state-of-the-art models,” the study stated.
The study disclosed that hallucinated content is also entering into government policy reports, court filings and scientific publications, raising doubts over regulation and appropriate use of AI tools.
“This growing concern highlights the need to audit the prevalence of hallucinations “in the wild”. While AI labs study hallucinations in simulated or controlled settings, we know less about their prevalence in deployed use.” the study noted.
The researchers stressed that systematic understanding of real-world hallucinations is important for the technical community, which develops foundation models, as well as for institutions, which rely on and vet knowledge work, such as courts, the patent office, media outlets, medical bodies, and peer review but at present there is not such evidence.
The research further explained hallucinations are difficult to study at scale because they are often embedded in long, unstructured text. In order to identify them, large samples of claims need to be extracted, retrieving factual evidence from external sources, and assessing the validity of each claim against a ground truth.
The study informed that human-based evaluations to measure hallucinations are costly to scale while automated LLM-based pipelines may reintroduce model-specific biases into the analysis.
Also Read: Half Right, Half Risky: AI Chatbots Wrong Half the Time on Health Advice
