{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,8,4]],"date-time":"2026-08-04T20:48:11Z","timestamp":1785876491133,"version":"3.56.0"},"reference-count":36,"publisher":"JMIR Publications Inc.","content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":["JMIR AI"],"abstract":"Abstract<\/jats:title>\n \n Background<\/jats:title>\n Colorectal cancer is a leading cause of cancer-related deaths in the United States, and colonoscopy remains the gold standard for early detection and prevention. However, many procedures are postponed due to inadequate bowel preparation, a preventable failure often caused by patients\u2019 difficulty in understanding and following written prep instructions. Prior interventions such as reminder apps and instructional videos have improved adherence only modestly, largely because they cannot answer patient-specific questions. Recent advances in large language models (LLMs) raise the possibility of developing conversational assistants that can provide interactive support to patients in procedure preparation.<\/jats:p>\n <\/jats:sec>\n \n Objective<\/jats:title>\n This study evaluated the correctness, harmfulness, and diversity of synthetic dialogues generated by leading LLMs acting as both simulated AI Coaches and patients for colonoscopy preparation.<\/jats:p>\n <\/jats:sec>\n \n Methods<\/jats:title>\n Five leading LLMs\u2014OpenAI\u2019s o3, GPT-4.1, and GPT-5.1; Meta\u2019s Llama 3.3 70B; and Mistral\u2019s Large-2411\u2014were used to generate 250 patient-AI Coach dialogues per model. Dialogues consisted of 3 to 7 question-answer pairs concerning diet, medications, and other prep-related topics. A multiprompt, multiquestion approach was designed to elicit diverse patient questions, and an error taxonomy was established to assess model capabilities in responding to questions. Human raters, including 3 medical experts, evaluated the generated questions for difficulty and the responses for correctness, error type, and potential harmfulness. Automatic evaluation using an LLM-as-a-judge approach complemented human evaluation. Question diversity was assessed using lexical diversity metrics (Distinct-1 and Distinct-2) and entropy. In addition, we evaluated a safety filtering mechanism in which responses judged incorrect by an automated evaluator were replaced with a deferral message instructing patients to contact their health care provider. Differences in response correctness across models were evaluated using permutation tests conducted at the dialogue level. Interrater agreement among human evaluators was assessed using the Gwet AC1 statistic. The study was conducted between May and September 2025.<\/jats:p>\n <\/jats:sec>\n \n Results<\/jats:title>\n Automatic evaluation results closely aligned with human judgments: leading models approached but did not achieve adequate performance. Closed-weight models (GPT-5.1, GPT-4.1, and o3) outperformed open-weight models (Llama and Mistral) on correctness, with the reasoning models (GPT-5.1 and o3) performing best. This turn-level ranking was preserved under the supplementary single-prompt baseline, although dialogue-level rankings differed. All models produced harmful errors, primarily due to omissions or misinterpretations of prep instructions. The multiprompt generation strategy substantially increased the diversity of patient questions compared with a single-prompt baseline. Applying an automated safety filter reduced overall error rates but failed to eliminate harmful responses.<\/jats:p>\n <\/jats:sec>\n \n Conclusions<\/jats:title>\n Although LLMs demonstrate strong potential to support colonoscopy preparation, none are yet reliable enough for unsupervised deployment in patient-facing contexts. Persistent harmful errors and the limited effectiveness of simple filtering mechanisms highlight the need for improved instruction adherence, stronger safety mechanisms, and validation using real patient queries.<\/jats:p>\n <\/jats:sec>","DOI":"10.2196\/88581","type":"journal-article","created":{"date-parts":[[2026,8,4]],"date-time":"2026-08-04T20:12:52Z","timestamp":1785874372000},"page":"e88581-e88581","source":"Crossref","is-referenced-by-count":0,"title":["Correctness, Harmfulness, and Diversity of Large Language Models for Colonoscopy Preparation Assistance: Comparative Evaluation Study"],"prefix":"10.2196","volume":"5","author":[{"ORCID":"https:\/\/orcid.org\/0009-0008-4737-7194","authenticated-orcid":false,"given":"Tomiris","family":"Kaumenova","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0009-0004-0524-022X","authenticated-orcid":false,"given":"Subhankar","family":"Chakraborty","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0001-8004-5169","authenticated-orcid":false,"given":"Eric","family":"Fosler-Lussier","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0009-0005-8721-1445","authenticated-orcid":false,"given":"Kebire","family":"Gofar","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0009-0004-0761-976X","authenticated-orcid":false,"given":"Isaiah","family":"Metcalf","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-5062-7958","authenticated-orcid":false,"given":"Andrew","family":"Perrault","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-3062-1719","authenticated-orcid":false,"given":"Michael","family":"White","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"1010","published-online":{"date-parts":[[2026,8,4]]},"reference":[{"key":"R1","unstructured":"Colorectal cancer statistics. Centers for Disease Control (CDC). URL: https:\/\/www.cdc.gov\/colorectal-cancer\/statistics\/index.html [Accessed 10-03-2026]"},{"key":"R2","doi-asserted-by":"publisher","DOI":"10.5888\/pcd19.220001","article-title":"Adults who have never been screened for colorectal cancer, Behavioral Risk Factor Surveillance System, 2012 and 2020","volume":"19","author":"Richardson","journal-title":"Prev Chronic Dis"},{"issue":"8","key":"R3","doi-asserted-by":"publisher","DOI":"10.2196\/26703","article-title":"Effects of a personalized smartphone app on bowel preparation quality: randomized controlled trial","volume":"9","author":"van der Zander","journal-title":"JMIR mHealth uHealth"},{"issue":"6","key":"R4","doi-asserted-by":"publisher","first-page":"942","DOI":"10.1177\/1090198119869964","article-title":"An automated text message navigation program improves the show rate for outpatient colonoscopy","volume":"46","author":"Mahmud","journal-title":"Health Educ Behav"},{"issue":"1","key":"R5","doi-asserted-by":"publisher","DOI":"10.1038\/ctg.2016.65","article-title":"A customized mobile application in colonoscopy preparation: a randomized controlled trial","volume":"8","author":"Sharara","journal-title":"Clin Transl Gastroenterol"},{"issue":"1","key":"R6","doi-asserted-by":"publisher","DOI":"10.1001\/jamanetworkopen.2020.34553","article-title":"Effect of text messaging on bowel preparation and appointment attendance for outpatient colonoscopy: a randomized clinical trial","volume":"4","author":"Mahmud","journal-title":"JAMA Netw Open"},{"issue":"1","key":"R7","doi-asserted-by":"publisher","DOI":"10.1001\/jamanetworkopen.2020.35720","article-title":"Texting to improve colonoscopy preparation and adherence needs more study","volume":"4","author":"Clancy","journal-title":"JAMA Netw Open"},{"issue":"3","key":"R8","doi-asserted-by":"publisher","first-page":"943","DOI":"10.1038\/s41591-024-03423-7","article-title":"Toward expert-level medical question answering with large language models","volume":"31","author":"Singhal","journal-title":"Nat Med"},{"key":"R9","unstructured":"Arora RK Wei J Hicks RS et al. HealthBench: evaluating large language models towards improved human health. arXiv. Preprint posted online on May 13, 2025. [doi: 10.48550\/arXiv.2505.08775]"},{"issue":"14","key":"R10","doi-asserted-by":"publisher","first-page":"6421","DOI":"10.3390\/app11146421","article-title":"What disease does this patient have? A large-scale open domain question answering dataset from medical exams","volume":"11","author":"Jin","journal-title":"Appl Sci"},{"key":"R11","doi-asserted-by":"crossref","unstructured":"Arya P Bloomquist M Chakraborty S et al. Bootstrapping a conversational guide for colonoscopy prep. In: Stoyanchev S Joty S Schlangen D Dusek O Kennington C Alikhani M , editors. Presented at: 24th Meeting of the Special Interest Group on Discourse and Dialogue; Sep 11-15, 2023:413-420; Prague, Czechia. [doi: 10.18653\/v1\/2023.sigdial-1.38]","DOI":"10.18653\/v1\/2023.sigdial-1.38"},{"key":"R12","doi-asserted-by":"publisher","DOI":"10.2196\/51300","article-title":"An AI dietitian for type 2 diabetes mellitus management based on large language and image recognition models: preclinical concept validation study","volume":"25","author":"Sun","journal-title":"J Med Internet Res"},{"key":"R13","doi-asserted-by":"publisher","DOI":"10.2196\/49240","article-title":"Clinical accuracy of large language models and Google Search responses to postpartum depression questions: cross-sectional study","volume":"25","author":"Sezgin","journal-title":"J Med Internet Res"},{"issue":"4","key":"R14","doi-asserted-by":"publisher","DOI":"10.1001\/jamanetworkopen.2024.4630","article-title":"Quality of large language model responses to radiation oncology patient care questions","volume":"7","author":"Yalamanchili","journal-title":"JAMA Netw Open"},{"issue":"6","key":"R15","doi-asserted-by":"publisher","first-page":"e379","DOI":"10.1016\/S2589-7500(24)00060-8","article-title":"The effect of using a large language model to respond to patient messages","volume":"6","author":"Chen","journal-title":"Lancet Digit Health"},{"issue":"6","key":"R16","doi-asserted-by":"publisher","first-page":"589","DOI":"10.1001\/jamainternmed.2023.1838","article-title":"Comparing physician and artificial intelligence chatbot responses to patient questions posted to a public social media forum","volume":"183","author":"Ayers","journal-title":"JAMA Intern Med"},{"issue":"10","key":"R17","doi-asserted-by":"publisher","DOI":"10.1001\/jamanetworkopen.2024.38535","article-title":"Perspectives on artificial intelligence-generated responses to patient messages","volume":"7","author":"Kim","journal-title":"JAMA Netw Open"},{"issue":"10","key":"R18","doi-asserted-by":"publisher","DOI":"10.1001\/jamanetworkopen.2023.36483","article-title":"Accuracy and reliability of chatbot responses to physician questions","volume":"6","author":"Goodman","journal-title":"JAMA Netw Open"},{"issue":"1","key":"R19","doi-asserted-by":"publisher","first-page":"77","DOI":"10.1038\/s41591-024-03328-5","article-title":"An evaluation framework for clinical use of large language models in patient interaction tasks","volume":"31","author":"Johri","journal-title":"Nat Med"},{"key":"R20","doi-asserted-by":"crossref","unstructured":"Li Z Zhu H Lu Z Yin M . Synthetic data generation with large language models for text classification: potential and limitations. Presented at: 2023 Conference on Empirical Methods in Natural Language Processing; Dec 6-10, 2023:10443-10461; Singapore. [doi: 10.18653\/v1\/2023.emnlp-main.647]","DOI":"10.18653\/v1\/2023.emnlp-main.647"},{"key":"R21","doi-asserted-by":"publisher","DOI":"10.1016\/j.smhl.2025.100570","article-title":"HealthQ: Unveiling questioning capabilities of LLM chains in healthcare conversations","volume":"36","author":"Wang","journal-title":"Smart Health"},{"key":"R22","doi-asserted-by":"crossref","unstructured":"Wei J Wang X Schuurmans D et al. Chain-of-thought prompting elicits reasoning in large language models. Presented at: Advances in Neural Information Processing Systems 35; Nov 28 to Dec 9, 2022:24824-24837; New Orleans, Louisiana, USA. [doi: 10.52202\/068431-1800]","DOI":"10.52202\/068431-1800"},{"key":"R23","unstructured":"1 day bowel prep with Miralax and Dulcolax. The Ohio State University Wexner Medical Center. URL: https:\/\/www.healthwise.net\/osumychart\/Content\/StdDocument.aspx?DOCHWID=custom.hs0146 [Accessed 06-06-2025]"},{"key":"R24","unstructured":"Zhang J Yu S Chong D et al. Verbalized sampling: how to mitigate mode collapse and unlock LLM diversity. arXiv. Preprint posted online on Oct 1, 2025. [doi: 10.48550\/arXiv.2510.01171]"},{"key":"R25","doi-asserted-by":"crossref","unstructured":"Xu W Zhu G Zhao X Pan L Li L Wang W . Pride and prejudice: LLM amplifies self-bias in self-refinement. Presented at: 62nd Annual Meeting of the Association for Computational Linguistics; Aug 11-16, 2024:15474-15492; Bangkok, Thailand. [doi: 10.18653\/v1\/2024.acl-long.826]","DOI":"10.18653\/v1\/2024.acl-long.826"},{"key":"R26","unstructured":"ATLAS.ti Team. Krippendorff\u2019s alpha: sample size and decision rules. ATLAS.ti 26 Windows User Manual. Lumivero; 2025. URL: https:\/\/manuals.atlasti.com\/Win\/en\/manual\/ICA\/ICASampleSizeAndDecisionRules.html [Accessed 07-07-2026]"},{"key":"R27","unstructured":"Gwet KL . Handbook of Inter-Rater Reliability, 4th Edition: The Definitive Guide to Measuring the Extent of Agreement Among Raters. 4th ed. Advanced Analytics; 2014. ISBN: 9780970806284"},{"key":"R28","doi-asserted-by":"crossref","unstructured":"Li D Jiang B Huang L et al. From generation to judgment: opportunities and challenges of LLM-as-a-judge. Presented at: 2025 Conference on Empirical Methods in Natural Language Processing; Nov 4-9, 2025:2757-2791; Suzhou, China. [doi: 10.18653\/v1\/2025.emnlp-main.138]","DOI":"10.18653\/v1\/2025.emnlp-main.138"},{"key":"R29","doi-asserted-by":"crossref","unstructured":"Papineni K Roukos S Ward T Zhu WJ . BLEU: a method for automatic evaluation of machine translation. Presented at: 40th Annual Meeting of the Association for Computational Linguistics; Jul 7-12, 2002:311-318; [doi: 10.3115\/1073083.1073135]","DOI":"10.3115\/1073083.1073135"},{"issue":"3","key":"R30","doi-asserted-by":"publisher","first-page":"393","DOI":"10.1162\/coli_a_00322","article-title":"A structured review of the validity of BLEU","volume":"44","author":"Reiter","journal-title":"Comput Linguist"},{"key":"R31","unstructured":"Wang T Kulikov I Golovneva O et al. Self-taught evaluators. arXiv. Preprint posted online on Aug 5, 2024. [doi: 10.48550\/arXiv.2408.02666]"},{"key":"R32","unstructured":"Lewis A White M Liu J et al. Winning big with small models: knowledge distillation vs. self-training for reducing hallucination in QA agents. Presented at: Fourth Workshop on Generation, Evaluation and Metrics (GEM). 705-727; URL: https:\/\/aclanthology.org\/2025.gem-1.62.pdf [Accessed 28-07-2026]"},{"key":"R33","unstructured":"Gulcehre C Paine TL Srinivasan S et al. Reinforced self-training (ReST) for language modeling. arXiv. Preprint posted online on Aug 17, 2023. [doi: 10.48550\/arXiv.2308.08998]"},{"key":"R34","unstructured":"Song Y Zhang H Eisenach C Kakade SM Foster D Ghai U . Mind the gap: examining the self-improvement capabilities of large language models. Presented at: Proceedings of the 13th International Conference on Learning Representations; Apr 24-28, 2025; Singapore. URL: https:\/\/proceedings.iclr.cc\/paper_files\/paper\/2025\/file\/63943ee9fe347f3d95892cf87d9a42e6-Paper-Conference.pdf [Accessed 07-07-2026]"},{"key":"R35","unstructured":"Eisenstein J Aghajani R Fisch A et al. Don\u2019t lie to your friends: learning what you know from collaborative self-play. Presented at: 1st Conference on Language Modeling (COLM 2025); Jul 8-11, 2025; Montr\u00e9al, Qu\u00e9bec, Canada. URL: https:\/\/openreview.net\/pdf?id=2vDJiGUfhV [Accessed 07-07-2026]"},{"key":"R36","unstructured":"Evaluating LLMs for colonoscopy preparation assistance. GitHub. 2025. URL: https:\/\/github.com\/sirimott\/prep-coach-llms [Accessed 20-11-2025]"}],"container-title":["JMIR AI"],"original-title":[],"language":"en","deposited":{"date-parts":[[2026,8,4]],"date-time":"2026-08-04T20:13:04Z","timestamp":1785874384000},"score":1,"resource":{"primary":{"URL":"https:\/\/ai.jmir.org\/2026\/1\/e88581"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,8,4]]},"references-count":36,"URL":"https:\/\/doi.org\/10.2196\/88581","relation":{},"ISSN":["2817-1705"],"issn-type":[{"value":"2817-1705","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,8,4]]},"article-number":"v5i8e88581"}}