Performance of Large Language Models in Oral Health Consultations and the Consistency of the 'AI-as-a-Judge' Framework.

OBJECTIVE

To evaluate the performance of large language models (LLMs) in responding to oral health consultations and to examine the consistency between the AI-as-a-Judge evaluation framework and human expert ratings.

METHODS

Nine oral health questions were selected from the World Dental Federation (FDI) official website and posed to 6 models: GPT-5.0, Gemini-3.0, DeepSeek-V3, Qwen3-Max, Kimi-K2 and Doubao-1.8-Pro. Responses were independently scored by 2 clinicians and 3 AI judges.

RESULTS

Significant performance differences were observed among the 6 models, with DeepSeek-V3 and Doubao-1.8-Pro achieving the best results. Inter-rater consistency among human experts was good (ICC = 0.860), while consistency among AI judges was low (ICC = 0.538). Human-AI consistency was extremely low (ICC = 0.215) and AI judges exhibited a significantly stricter scoring tendency.

CONCLUSION

Leading domestic LLMs have attained competitive performance in oral health consultations. However, the current 'AI-as-a-Judge' framework demonstrates significant inconsistency and bias compared to human experts, suggesting that automated AI evaluation systems are not yet a reliable substitute for human expert review in clinical contexts.