{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,8,5]],"date-time":"2026-08-05T05:52:13Z","timestamp":1785909133401,"version":"3.56.0"},"reference-count":39,"publisher":"Frontiers Media SA","license":[{"start":{"date-parts":[[2026,8,5]],"date-time":"2026-08-05T00:00:00Z","timestamp":1785888000000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":["frontiersin.org"],"crossmark-restriction":true},"short-container-title":["Front. Public Health"],"abstract":"<jats:sec>\n                    <jats:title>Introduction<\/jats:title>\n                    <jats:p>Large language models (LLMs) are entering drug-safety and regulatory workflows, yet their behavior at the boundary between unavailable evidence and risk reassurance remains poorly characterized. Drug-induced liver injury (DILI) is a stringent setting because evidence is fragmented across labels, case reports, mechanistic studies, and curated knowledge bases, while unsupported low-risk reassurance can be consequential.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Methods<\/jats:title>\n                    <jats:p>We evaluated five LLMs on a fixed 47-drug DILI risk-assessment panel using 1,410 parsed responses from closed-book answering and an evidence-gated protocol that restricted responses to supplied PubMed-derived evidence. Curated DILI resources were excluded from prompts and used only for evaluation. After filtering, 38 drugs had no direct DILI decision-support evidence in the supplied packet, and 9 had direct DILI-relevant evidence; a post hoc PubMed title\/abstract recall stress test identified five additional drugs with recoverable direct DILI evidence outside the packet.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Results<\/jats:title>\n                    <jats:p>In the no-direct-evidence slice defined by the supplied packet, closed-book models rarely abstained, with drug-level abstention ranging from 5.3% to 33.3%; the evidence-gated protocol required abstention, which all models followed for every no-direct-evidence drug. The same pattern held for recent or low-recognition drugs, where evidence-gated abstention reached 92.0% to 100.0% vs. 8.0% to 49.3% under closed-book answering. Closed-book models also produced high-confidence low-risk responses for DILI-positive drugs, a label-discordant pattern largely removed by evidence gating. Independent expert review of selected responses showed that label discordance did not always imply a clinically unreasonable low-risk category, but identified unsafe reassurance through overconfident wording and under-cautious responses in selected cases. When direct DILI evidence was provided, all models preserved citation-grounded non-abstaining answers. However, they differed in how often they committed to a conclusive rather than an uncertain risk category. Citation-bearing evidence-gated responses cited only the supplied PubMed identifiers and achieved 91.2% to 100.0% concordance with the supplied grade.<\/jats:p>\n                  <\/jats:sec>\n                  <jats:sec>\n                    <jats:title>Discussion<\/jats:title>\n                    <jats:p>These findings identify unsupported reassurance as measurable evidence-boundary behavior in LLM drug-risk assessment and establish a reproducible framework for auditing adherence to an externally supplied evidence boundary, defined by PubMed evidence classification and enforced by prompt policy rather than inferred independently by the model.<\/jats:p>\n                  <\/jats:sec>","DOI":"10.3389\/fpubh.2026.1904062","type":"journal-article","created":{"date-parts":[[2026,8,5]],"date-time":"2026-08-05T05:39:03Z","timestamp":1785908343000},"update-policy":"https:\/\/doi.org\/10.3389\/crossmark-policy","source":"Crossref","is-referenced-by-count":0,"title":["Confident but unsupported: auditing large language models against supplied evidence boundaries in drug-induced liver injury assessment"],"prefix":"10.3389","volume":"14","author":[{"given":"Jinwan","family":"Shi","sequence":"first","affiliation":[{"name":"Foshan Sanshui District People's Hospital","place":["Foshan, Guangdong, China"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Yong","family":"Ma","sequence":"additional","affiliation":[{"name":"School of Investigation, Zhengzhou Police University","place":["Zhengzhou, Henan, China"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Yinhui","family":"Liu","sequence":"additional","affiliation":[{"name":"Foshan Sanshui District People's Hospital","place":["Foshan, Guangdong, China"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Bowen","family":"Shi","sequence":"additional","affiliation":[{"name":"School of Medical Information Engineering, Guangdong Pharmaceutical University","place":["Guangzhou, Guangdong, China"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"1965","published-online":{"date-parts":[[2026,8,5]]},"reference":[{"key":"B1","doi-asserted-by":"publisher","first-page":"259","DOI":"10.1038\/s41586-023-05881-4","article-title":"Foundation models for generalist medical artificial intelligence","volume":"616","author":"Moor","year":"2023","journal-title":"Nature"},{"key":"B2","doi-asserted-by":"publisher","first-page":"172","DOI":"10.1038\/s41586-023-06291-2","article-title":"Large language models encode clinical knowledge","volume":"620","author":"Singhal","year":"2023","journal-title":"Nature"},{"key":"B3","doi-asserted-by":"publisher","first-page":"1930","DOI":"10.1038\/s41591-023-02448-8","article-title":"Large language models in medicine","volume":"29","author":"Thirunavukarasu","year":"2023","journal-title":"Nat Med"},{"key":"B4","doi-asserted-by":"publisher","first-page":"221","DOI":"10.1038\/s41746-026-02353-7","article-title":"Scaling medical device regulatory science using large language models","volume":"9","author":"Li","year":"2026","journal-title":"npj Digit Med"},{"key":"B5","doi-asserted-by":"publisher","first-page":"2613","DOI":"10.1038\/s41591-024-03097-1","article-title":"Evaluation and mitigation of the limitations of large language models in clinical decision-making","volume":"30","author":"Hager","year":"2024","journal-title":"Nat Med"},{"key":"B6","doi-asserted-by":"publisher","first-page":"367","DOI":"10.1038\/s41746-024-01366-4","article-title":"Probabilistic medical predictions of large language models","volume":"7","author":"Gu","year":"2024","journal-title":"npj Digit Med"},{"key":"B7","doi-asserted-by":"publisher","first-page":"618","DOI":"10.1038\/s41591-024-03445-1","article-title":"Medical large language models are vulnerable to data-poisoning attacks","volume":"31","author":"Alber","year":"2025","journal-title":"Nat Med"},{"key":"B8","doi-asserted-by":"publisher","first-page":"274","DOI":"10.1038\/s41746-025-01670-7","article-title":"A framework to assess clinical safety and hallucination rates of large language models for medical text summarisation","volume":"8","author":"Asgari","year":"2025","journal-title":"npj Digit Med"},{"key":"B9","doi-asserted-by":"publisher","first-page":"177","DOI":"10.1038\/s41746-024-01180-y","article-title":"The long but necessary road to responsible use of large language models in healthcare research","volume":"7","author":"Kwong","year":"2024","journal-title":"npj Digit Med"},{"key":"B10","doi-asserted-by":"publisher","first-page":"1706383","DOI":"10.3389\/fdgth.2025.1706383","article-title":"Epistemic and ethical limits of large language models in evidence-based medicine: from knowledge to judgment","volume":"7","author":"Qi","year":"2026","journal-title":"Front Digit Health"},{"key":"B11","doi-asserted-by":"publisher","first-page":"1379260","DOI":"10.3389\/fdsfr.2024.1379260","article-title":"Assessing the performance of large language models in literature screening for pharmacovigilance: a comparative study","volume":"4","author":"Li","year":"2024","journal-title":"Front Drug Saf Regul"},{"key":"B12","doi-asserted-by":"publisher","first-page":"27886","DOI":"10.1038\/s41598-025-09138-0","article-title":"The need for guardrails with large language models in pharmacovigilance and other medical safety critical settings","volume":"15","author":"Hakim","year":"2025","journal-title":"Sci Rep"},{"key":"B13","doi-asserted-by":"publisher","first-page":"1392180","DOI":"10.3389\/fpubh.2024.1392180","article-title":"BERT-based language model for accurate drug adverse event extraction from social media: implementation, evaluation, and contributions to pharmacovigilance practices","volume":"12","author":"Dong","year":"2024","journal-title":"Front Public Health"},{"key":"B14","doi-asserted-by":"publisher","DOI":"10.1080\/10543406.2024.2403442.","article-title":"Developing large language models to detect adverse drug events in posts on X","author":"Deng","year":"2024","journal-title":"J Biopharm Stat"},{"key":"B15","doi-asserted-by":"publisher","first-page":"20420986251339358","DOI":"10.1177\/20420986251339358","article-title":"Can large language models detect drug-drug interactions leading to adverse drug reactions?","volume":"16","author":"Sicard","year":"2025","journal-title":"Ther Adv Drug Saf"},{"key":"B16","doi-asserted-by":"publisher","first-page":"5490","DOI":"10.3390\/jcm14155490","article-title":"Large Language models for adverse drug events: a clinical perspective","volume":"14","author":"Zitu","year":"2025","journal-title":"J Clin Med"},{"key":"B17","first-page":"9459","article-title":"Retrieval-augmented generation for knowledge-intensive NLP tasks","volume":"33","author":"Lewis","year":"2020","journal-title":"Adv Neural Inf Process Syst"},{"key":"B18","doi-asserted-by":"publisher","first-page":"i119","DOI":"10.1093\/bioinformatics\/btae238","article-title":"Improving medical reasoning through retrieval and self-reflection with retrieval-augmented large language models","volume":"40","author":"Jeong","year":"2024","journal-title":"Bioinformatics"},{"key":"B19","doi-asserted-by":"publisher","first-page":"1635381","DOI":"10.3389\/fpubh.2025.1635381","article-title":"MEGA-RAG: a retrieval-augmented generation framework with multi-evidence guided answer refinement for mitigating hallucinations of LLMs in public health","volume":"13","author":"Xu","year":"2025","journal-title":"Front Public Health"},{"key":"B20","doi-asserted-by":"publisher","first-page":"20","DOI":"10.1038\/s41746-024-01010-1","article-title":"Diagnostic reasoning prompts reveal the potential for large language model interpretability in medicine","volume":"7","author":"Savage","year":"2024","journal-title":"npj Digit Med"},{"key":"B21","doi-asserted-by":"publisher","first-page":"1","DOI":"10.1145\/3571730","article-title":"Survey of hallucination in natural language generation","volume":"55","author":"Ji","year":"2023","journal-title":"ACM Comput Surv"},{"key":"B22","doi-asserted-by":"publisher","first-page":"288","DOI":"10.1038\/s41746-024-01282-7","article-title":"Medical large language models are susceptible to targeted misinformation attacks","volume":"7","author":"Han","year":"2024","journal-title":"npj Digit Med"},{"key":"B23","doi-asserted-by":"publisher","first-page":"58","DOI":"10.1038\/s41572-019-0105-0","article-title":"Drug-induced liver injury","volume":"5","author":"Andrade","year":"2019","journal-title":"Nat Rev Dis Primers"},{"key":"B24","doi-asserted-by":"publisher","first-page":"1222","DOI":"10.1016\/j.jhep.2019.02.014","article-title":"EASL clinical practice guidelines: drug-induced liver injury","volume":"70","year":"2019","journal-title":"J Hepatol"},{"key":"B25","doi-asserted-by":"publisher","first-page":"1036","DOI":"10.1002\/hep.32689","article-title":"AASLD practice guidance on drug, herbal, and dietary supplement-induced liver injury","volume":"77","author":"Fontana","year":"2023","journal-title":"Hepatology"},{"key":"B26","doi-asserted-by":"crossref","unstructured":"Drug-Induced Liver Injury: Premarketing Clinical Evaluation\n          \n          \n          2009","DOI":"10.69645\/UFJZ3421"},{"key":"B27","unstructured":"Drug-Induced Liver Injury (DILI): Current Status and Future Directions for Drug Development and the Post-market Setting\n          \n          \n          2020"},{"key":"B28","doi-asserted-by":"publisher","first-page":"104485","DOI":"10.1016\/j.drudis.2025.104485","article-title":"DILIrank 20: an updated and expanded database for drug-induced liver injury risk based on FDA labeling and a literature review","volume":"30","author":"Olubamiwa","year":"2025","journal-title":"Drug Discov Today"},{"key":"B29","unstructured":"Drug Induced Liver Injury Rank (DILIrank 2\n          \n          2025"},{"key":"B30","doi-asserted-by":"publisher","first-page":"201","DOI":"10.1016\/j.drudis.2019.09.022","article-title":"Drug-induced liver injury severity and toxicity (DILIst): binary classification of 1279 drugs by human hepatotoxicity","volume":"25","author":"Thakkar","year":"2020","journal-title":"Drug Discov Today"},{"key":"B31","unstructured":"Drug-Induced Liver Injury Severity and Toxicity (DILIst) Dataset\n          \n          \n          2020"},{"key":"B32","unstructured":"LiverTox: Clinical and Research Information on Drug-Induced Liver Injury\n          \n          \n          2012"},{"key":"B33","doi-asserted-by":"publisher","first-page":"e0000198","DOI":"10.1371\/journal.pdig.0000198","article-title":"Performance of ChatGPT on USMLE: potential for AI-assisted medical education using large language models","volume":"2","author":"Kung","year":"2023"},{"key":"B34","doi-asserted-by":"publisher","first-page":"263","DOI":"10.1038\/s41746-025-01684-1","article-title":"Evaluating large language model workflows in clinical decision support for triage and referral and diagnosis","volume":"8","author":"Gaber","year":"2025","journal-title":"npj Digit Med"},{"key":"B35","doi-asserted-by":"publisher","first-page":"101579","DOI":"10.1016\/j.jhepr.2025.101579","article-title":"Evaluating large language models' performance in answering common questions on drug-induced liver injury","volume":"7","author":"Dong","year":"2025","journal-title":"JHEP Rep"},{"key":"B36","doi-asserted-by":"publisher","first-page":"1524","DOI":"10.1021\/acs.chemrestox.4c00134","article-title":"Toward an explainable large language model for the automatic identification of the drug-induced liver injury literature","volume":"37","author":"Ma","year":"2024","journal-title":"Chem Res Toxicol"},{"key":"B37","doi-asserted-by":"publisher","first-page":"102627","DOI":"10.1016\/j.jceh.2025.102627","article-title":"Toward real-time detection of drug-induced liver injury using large language models: a feasibility study from clinical notes","volume":"15","author":"Suenghataiphorn","year":"2025","journal-title":"J Clin Exp Hepatol"},{"key":"B38","unstructured":"Compilation of CDER New Molecular Entity (NME) Drug and New Biologic Approvals\n          \n          \n          2026"},{"key":"B39","doi-asserted-by":"publisher","first-page":"60","DOI":"10.1038\/s41591-024-03425-5","article-title":"The TRIPOD-LLM reporting guideline for studies using large language models","volume":"31","author":"Gallifant","year":"2025","journal-title":"Nat Med"}],"container-title":["Frontiers in Public Health"],"original-title":[],"link":[{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fpubh.2026.1904062\/full","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,8,5]],"date-time":"2026-08-05T05:39:06Z","timestamp":1785908346000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fpubh.2026.1904062\/full"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,8,5]]},"references-count":39,"alternative-id":["10.3389\/fpubh.2026.1904062"],"URL":"https:\/\/doi.org\/10.3389\/fpubh.2026.1904062","relation":{},"ISSN":["2296-2565"],"issn-type":[{"value":"2296-2565","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,8,5]]},"article-number":"1904062"}}