In an exploratory blinded comparative study evaluating informed consent documentation for targeted therapies, large language model generated texts achieved higher overall content quality scores than physician drafted forms but demonstrated higher reading levels and a greater frequency of hallucinations.
The investigation examined documentation for targeted therapies in rheumatology using a single-center, retrospective, paired-case, blinded multi-evaluator design conducted at Beijing Tongren Hospital. Researchers enrolled 70 adult patients who received targeted therapy in the hospital’s Department of Rheumatology during 2025.
For each case, large language model texts were generated using DeepSeek-V4-pro based on identical de-identified clinical case data and standardized prompts. Corresponding physician-drafted consent texts were extracted directly from existing electronic health records, yielding a total of 140 texts for comparative analysis.
Content quality across basic information, risks, benefits, and alternative treatment options was evaluated by three rheumatologists performing blinded scoring using a locally developed assessment framework comprising 8 domains and 13 items. Informational errors were identified through source tracing, and reading level was quantified using CRIE 3.0.
The language model arm achieved a median overall quality score of 60 compared to 50 for the physician arm, reflecting an adjusted mean difference of 9.95 points with a 95 percent confidence interval of 9.36 to 10.54. Evaluator scores favored the model across the specific domains of risks, benefits, and alternative options.
However, the model-generated documents exhibited significantly higher reading level scores, averaging 7.87 compared to 7.10 in the physician arm. The model arm also demonstrated a 27.1 percent rate of hallucinations involving unsupported patient-specific inferences, whereas the physician arm had zero hallucinations.
No statistically significant differences between the two study arms were observed regarding factual inaccuracies or omissions. The authors concluded that while large language models more consistently covered predefined content domains and may assist in drafting structured consent content, they require physician review and item-by-item verification prior to clinical use.
Subject of Research: Comparison of large language model and physician-authored informed consent content for targeted therapies
Article Title: Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study
Article References: Xie, Y., Gao, Y., Peng, J., Tao, Y., Xu, S., Cui, Z., Zhou, B., Wang, E., & Cui, L. (2026). Quality, readability, and information fidelity of large language model–generated versus physician-authored informed-consent content for targeted therapies: a blinded, paired comparative study. Frontiers in Pharmacology, 17, Article 1958826. https://doi.org/10.3389/fphar.2026.1958826
Image Credits: AI Generated
DOI: 10.3389/fphar.2026.1958826
Keywords: informed consent, large language models, targeted therapy, rheumatology, clinical documentation, artificial intelligence in medicine, readability, hallucinations, Quality, information, fidelity, large
News Source: Ophelia Keating. (November 13, 2026). Study Evaluates AI and Physician Informed Consent for Targeted Therapies. Scienmag.



