Related Experiment Video
Updated: Sep 4, 2026
A Postoperative Evaluation Guideline for Computer-Assisted Reconstruction of the Mandible
Published on: January 28, 2020
Guideline-Based Evaluation of Five Generative Artificial Intelligence Chatbot Platforms for Clinician-Oriented
Selin Gaş1, Erdinç Sulukan2, Büşra Korkmaz1
1Department of Oral and Maxillofacial Surgery, Faculty of Dentistry, Istanbul Gelisim University, Cihangir Mah., Şehit Jandarma Komando Er Hakan Öner Sok. No: 1, 34310 Avcılar, Istanbul, Turkey.
Objective:
To compare the quality and readability of responses from five generative artificial intelligence chatbot platforms to clinician-oriented questions on open temporomandibular joint (TMJ) surgery against guideline-based reference answers.
Material And Methods:
Forty questions across eight domains were submitted on 16 July 2026 to ChatGPT (GPT-5.5), Claude (Opus 4.8), Gemini (3.1 Pro), Grok (4) and Perplexity (Pro), each via its paid tier at default settings (200 responses). Two blinded oral and maxillofacial surgeons applied the Quality Analysis of Medical Artificial Intelligence (QAMAI) tool, the Global Quality Score (GQS) and a five-point overall quality rating using a priori key elements and written anchors. Readability was assessed with the Flesch-Kincaid Grade Level (FKGL) and Flesch Reading Ease Score (FRES); platforms were compared with the Friedman test and Bonferroni-corrected Wilcoxon post-hoc tests; inter-rater reliability used intraclass correlation coefficients (ICC).
Results:
Inter-rater reliability was good to excellent (average-measure ICC 0.93-0.99). All outcomes except clarity differed among platforms (P < .001). Perplexity achieved the highest QAMAI total (27.5 ± 1.4; Kendall's W = 0.75), largely through retrieval-based source provision; excluding this domain, Claude, Perplexity and Gemini converged. Claude had the highest GQS (4.5 ± 0.6), Gemini the highest overall rating (4.3 ± 0.7); Grok scored lowest. Claude and Gemini were least readable (median FKGL 22.7 and 26.1; FRES -11.6 and -7.5). Length did not explain scores within platforms.
Conclusions:
Performance varied substantially across platforms and dimensions; no platform optimized all outcomes. These findings describe informational quality, not clinical safety or decision-making, and support specialist verification before clinical use.
