Multimodal AI chatbots in oral and maxillofacial radiology: added images help, misleading cues hurt

In 90 oral/maxillofacial cases, three multimodal artificial intelligence chatbots were not clinically usable: panoramic-only accuracy for diseased cases was 0-60%, rising to 38-92% with added info. Cropped images helped most; misleading cues lowered accuracy.

• Cropped panoramic images were the most consistently beneficial added visual input, while axial cone-beam CT images provided less improvement.
• Model behaviour differed: GPT-5.1 missed lesions, Gemini 3 Flash was sensitive but less specific, and Claude Opus 4.7 tended toward benign diagnoses.
• Gemini 3 Flash accepted a false benign suggestion in 96% of cases it had initially classified as normal; specific-diagnosis accuracy for benign lesions was about half of category-level accuracy.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.