The detectability paradox: bilingual medical report generation with open-weight models and the limits of human oversight

Article Article scientifique Article The detectability paradox: bilingual medical report generation with open-weight models and the limits of human oversight eng 2026-05-08 2026-07-01 author Rouhizadeh Hossein Department of Radiology and Medical Informatics, Faculty of Medicine, University of Geneva, Geneva, 1202, Switzerland. 986049 Département de radiologie et informatique médicale (RADIO) author Sandralegar Abiram 0009-0007-2724-6632 Division of Neurosurgery, Geneva University Hospitals, Geneva, 1211, Switzerland. 907767 Département des neurosciences cliniques (OTONO) author Yazdani Anthony 0000-0003-3309-6128 Department of Radiology and Medical Informatics, Faculty of Medicine, University of Geneva, Geneva, 1202, Switzerland. 838059 Département de radiologie et informatique médicale (RADIO) author Feng Weibo Department of Radiology and Medical Informatics, Faculty of Medicine, University of Geneva, Geneva, 1202, Switzerland. 431332 Département de radiologie et informatique médicale (RADIO) author Schreier Oren Laurent Department of Radiology and Medical Informatics, Faculty of Medicine, University of Geneva, Geneva, 1202, Switzerland. 923465 author Kim Yonnou Faculty of Medicine, University of Geneva, Geneva, 1211, Switzerland. 962541 Structure enseignement (ENSI) author Sirbal Assiya Faculty of Medicine, University of Geneva, Geneva, 1211, Switzerland. 937381 author Pirelli Valentino Faculty of Medicine, University of Geneva, Geneva, 1211, Switzerland. 400651 author Yang Rui Centre for Quantitative Medicine, Duke-NUS Medical School, Singapore, 169857, Singapore. author Sveikata Lukas 0000-0001-8015-7929 Division of Neurology, Department of Clinical Neurosciences, Geneva University Hospitals and Faculty of Medicine, Geneva, 1211, Switzerland. 759773 Département des neurosciences cliniques (OTONO) author Tessitore Elena Cardiology Division, Internal Medicine Department, Geneva University Hospitals, Geneva, 1211, Switzerland. 750991 Département de médecine (DEMED) author Liu Nan 0000-0003-3610-4883 Centre for Quantitative Medicine, Duke-NUS Medical School, Singapore, 169857, Singapore. author Bijlenga Philippe Alexandre Pierre 0000-0002-3586-2757 Division of Neurosurgery, Geneva University Hospitals, Geneva, 1211, Switzerland. 104899 Département des neurosciences cliniques (OTONO) author Teodoro Douglas 0000-0001-6238-4503 Department of Radiology and Medical Informatics, Faculty of Medicine, University of Geneva, Geneva, 1202, Switzerland. 682380 Département de radiologie et informatique médicale (RADIO) Douglas Henrique Teodoro 1303-1313 10.1093/jamia/ocag070 42097830 1067-5027 PMC13317965 https://academic.oup.com/jamia/article/33/7/1303/8672665 Objectives: The automation of medical report generation using large language models (LLMs) could significantly reduce physicians' documentation burden while enhancing healthcare efficiency. However, the misuse of generative artificial intelligence in medical reporting can lead to important safety risks for patients. We addressed 2 questions: (1) What is the quality of medical reports generated by LLMs in English and French? and (2) Can we distinguish between human-written and LLM-generated medical reports? Materials and methods: We evaluated the quality of reports generated by several multilingual, open-weight LLMs using text similarity metrics on 4212 medical reports in English and French across multiple specialties. A bilingual expert panel of certified physicians (n = 4) and medical residents (n = 5) scored accuracy, fluency, and completeness of generated reports using a 1-5 Likert scale. Experts also completed a Turing-like test, blindly identifying reports as human or machine-generated. Results: Phi-4 achieved the best overall performance (ROUGE-1: 0.70, BERTScore: 0.83). Expert evaluation confirmed high-quality reports in both languages (overall 4.6/5.0). Medical experts performed better than chance but struggled to differentiate human versus machine reports (accuracy: 0.60). Automatic classifiers showed strong performance (accuracy: 0.98). Discussion: The high quality of LLM-generated reports supports their potential to enhance healthcare efficiency in multilingual settings. However, the discrepancy between human detection difficulty and automated detection success reveals inherent limitations in relying solely on human oversight for quality assurance and misuse prevention. Conclusions: Deployment of LLMs for medical reporting requires combining automated detection tools with human expertise to ensure patient safety. Dataset and code: https://github.com/ds4dh/medical_report_generation. Automated report generation Electronic health records (EHRs) Large language models (LLMs) Medical documentation Multilingual medical reports Dewey 616.8 Dewey 616.0757 Dewey 616 62 Département de radiologie et informatique médicale 227 61 Département des neurosciences cliniques 226 56 Département de médecine 221 false 928 Neuropsychologie et neurologie comportementale 951 1379 DS4DH - Data Science for Digital Health 1035 513 Groupe Schaller Karl Lothard (neurochirurgie) 851 https://github.com/ds4dh/medical_report_generation Journal of the American Medical Informatics Association 33 7 6059dc95-e1b9-4c64-b4fc-5465714e729a Article (Published version) Public ocag070.pdf 2016960 application/pdf CC BY-4.0 c3adbd1a2884d167946eae447debb132afd72224cbd3cecf62615e42bb49d2ee ee06f432-cdf5-4784-8db8-6c279a92145a Supplemental data Public ocag070_supplementary_data.zip 15437039 application/zip b440f4b10b23bd26dc744302d28e2529c82d150e6d4c286ae5f4bbe5cde0bf0e openAccess Faculté de médecine / Section de médecine clinique / Département de radiologie et informatique médicale Faculté de médecine / Section de médecine clinique / Département des neurosciences cliniques Faculté de médecine / Section de médecine clinique / Département de médecine

Title	Published in	Access level	OA Policy	Year	Views	Downloads

The detectability paradox: bilingual medical report generation with open-weight models and the limits of human oversight	Journal of the American Medical Informatics Association			2026	4	1

Archive ouverte UNIGE

Kim, Yonnou