İZİM TÜRKER KADER, BURÇİN ARICAN
Journal of Health Sciences and Medicine - 2025;8(4):697-702
Aims: This study aimed to evaluate the readability of informed consent forms for provisional crowns and bridges by comparing a human-written version with AI-generated texts produced by two large language models (LLMs): GPT-4o (OpenAI) and Claude 3.7 Sonnet (Anthropic). Methods: A three-page informed consent form authored by a prosthodontic specialist was used as a human-written reference. Using identical structured prompts, comparable consent forms were generated by GPT-4o and Claude 3.7 Sonnet. Specifically, the models were instructed to first explain the clinical purpose of provisional dental restorations and then generate a three-page patient-oriented informed consent form, avoiding unnecessary technical jargon and adopting the tone of a prosthodontic specialist. The prompts guided the models to address each section sequentially, including: title of the form, patient identification, introductory statement, treatment and procedures, expected benefits, expected outcomes without treatment, treatment alternatives, possible risks and complications, estimated duration of the procedure, and signature section. Readability was assessed using the Flesch-Kincaid Grade Level (FKGL) metric, along with descriptive comparisons of word count, sentence count, and passive voice percentage. Results: The human-written form consisted of 1158 words, achieved an FKGL score of 10.8, and contained 34.5% passive voice. The GPT-4o form showed 956 words, an FKGL of 12.6, and 20.4% passive voice. The Claude 3.7 Sonnet form had 1338 words, an FKGL of 14.7, and 35% passive voice. These results revealed marked differences in document length, sentence count, and passive voice usage, with the AI-generated texts displaying more complex sentence structures and higher reading grade levels. Conclusion: Although all forms exceeded the recommended readability level for patient-facing documents, the AI-generated versions-particularly the Claude 3.7 Sonnet form-were more difficult to read due to greater length and more complex sentence structure. These results underscore the importance of human oversight in editing and simplifying AI-generated materials, ensuring they meet the readability standards essential for patient comprehension.
