Hallucination Rate of Peer-Reviewed Citations Generated by Large Language Models in Neurocritical Care
TL;DR
Under standardized zero-shot, retrieval-disabled web-interface conditions, LLMs generated substantial numbers of inaccurate and fabricated NCC citations, which should be verified across reliable databases before use in clinical, educational, or scholarly work.
Abstract
Importance
Large language models (LLMs) are increasingly used for scientific literature retrieval, yet their citation accuracy in specialized clinical domains remains poorly characterized. In neurocritical care (NCC), fabricated or inaccurate citations may be difficult to detect without deliberate verification.
Objectives
To evaluate hallucination and fabrication rates of peer-reviewed citations generated by three LLMs across core NCC topics, under constrained zero-shot, memory-only conditions. DESIGN, SETTING, AND
Participants
In this cross-sectional, blinded technology performance evaluation, Generative Pretrained Transformer (GPT)-5.3, DeepSeek-V3, and Grok-4 were queried on March 10, 2026, under identical zero-shot, retrieval-disabled web-interface conditions. Ten NCC topics were submitted to each model, and each model generated 10 references per topic, yielding 300 references. MAIN OUTCOMES AND MEASURES: Two NCC experts, blinded to model identity, independently verified each reference against PubMed, DOI, Google Scholar, and CrossRef and scored accuracy using a Hallucination Scale (0–3). The primary outcome was any hallucination, defined as any citation inaccuracy. The secondary outcome was fabrication, defined as a nonexisting complete bibliographic entity.
Results
Inter-rater agreement was excellent (κ = 0.91; 95% CI, 0.86–0.96). Overall, 165 of 300 references (55.0%) contained a citation inaccuracy, and 85 of 300 (28.3%) were completely fabricated. DeepSeek-V3 had the lowest hallucination rate (23%; fabrication 8%), followed by GPT-5.3 (69%; fabrication 27%) and Grok-4 (73%; fabrication 50%). Compared with DeepSeek-V3, Grok-4 was 3.17 times more likely to hallucinate (95% CI, 2.03–4.96; p < 0.001), and GPT-5.3 was 3.00 times more likely to hallucinate (95% CI, 1.94–4.63; p < 0.001). Topic-level findings were exploratory and should be interpreted cautiously.
Conclusions
AND RELEVANCE: Under standardized zero-shot, retrieval-disabled web-interface conditions, LLMs generated substantial numbers of inaccurate and fabricated NCC citations. Because fabricated references can appear complete and credible, artificial intelligence-generated citations should be verified across reliable databases before use in clinical, educational, or scholarly work.