Answers are grounded in this paper and cite their sources.
Introduction and aims To evaluate the authenticity detection performance of multimodal large language models (MLLMs) using real and AI-generated intraoral photographs and to compare their performance with human evaluators with different levels of dental experience. Methods A dataset of 1,000 intraoral photographs (500 real and 500 AI-generated) across 5 standardised views was constructed from single-institution clinical images and the publicly available AI-Hub dataset. GPT-5.1 (OpenAI), Claude Opus 4.5 (Anthropic), and Gemini 3.0 Pro (Google) evaluated all images through vendor APIs. Human evaluation was conducted using a stratified random subset of 200 images by fourth-year dental students, dentists with ≤5 years, and dentists with ≥6 years of clinical experience (n = 5 per group). Results In the 200-image subset, dentists with ≥6 years showed the highest performance (accuracy: 0.740; F1-score: 0.675), followed by dentists with ≤5 years (accuracy: 0.690; F1-score: 0.557). Across the 1,000-image dataset, GPT-5.1 achieved the highest performance among the MLLMs (accuracy: 0.646; F1-score: 0.452), whereas Claude Opus 4.5 showed the lowest (accuracy: 0.571; F1-score: 0.249). All groups demonstrated substantially higher accuracy for real images than for the curated, visually plausible AI-generated images, with the MLLMs combining near-perfect specificity (0.998-1.000) and low recall (0.142-0.292), whereas human specificity was 0.890-0.990 and recall 0.390-0.540. The mandibular occlusal view showed the lowest accuracy among the MLLM groups. Conclusion Current MLLMs showed lower performance than experienced human evaluator panels, with limited recall for the AI-generated images in this study, suggesting that they are not yet suitable as standalone authenticity detection tools in dentistry. Clinical Relevance The increasing visual realism of generative artificial intelligence-generated intraoral photographs suggests that authenticity cannot be judged from visual inspection alone, and robust validation is needed before such images are used as clinical records, teaching material, or training data for artificial intelligence.
No takes yet. Share an insight, caveat, or question.
Lee et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: