Related Experiment Videos
Evaluating Internet-Based AI Chatbots as Unregulated Digital Health Resources in Dental Emergency Management:
Ahmed Bashah1, Quan Huixin1, Hamza Zaawari2
1Department of Stomatology, The First Affiliated Hospital of Dalian Medical University, No.222, Zhongshan Road, Dalian, Liaoning, 116011, China, +86 18098873280.
Background:
Publicly accessible, internet-based AI chatbots have emerged as unregulated digital health resources that clinicians and patients consult in real time, including during acute medical emergencies. In dental practice-where life-threatening events such as anaphylaxis, cardiac arrest, and severe hypoglycemia can unfold within seconds-the unsupervised use of these tools as point-of-care decision aids poses uncharacterized patient safety risks. Despite widespread deployment, no study has systematically evaluated the accuracy, temporal consistency, or multilingual reliability of contemporary large language model (LLM)-based chatbots specifically for dental medical emergency management.
Objective:
This study aimed to evaluate the accuracy, temporal stability, and multilingual performance of five contemporary LLM-based chatbots when queried on the management of medical emergencies in dental practice.
Methods:
A comparative study was conducted to assess five AI chatbots: ChatGPT-5.2 (OpenAI), Gemini 3.1 (Google), DeepSeek-R1 (DeepSeek AI), Grok 4.20 (xAI), and Kimi K2.5 (Moonshot AI). A total of 130 questions covering dental medical emergency management were developed and categorized into four formats: multiple-choice questions (MCQs), yes-or-no questions, open-ended WH questions, and clinical scenario questions. Each question was presented using two prompting strategies (basic and advanced) across three languages (English, Arabic, and Chinese). Data collection was performed over seven consecutive days, generating 27,300 chatbot responses. Responses were evaluated by language-specific expert reviewers and classified as correct or incorrect based on established clinical references. Statistical analysis was performed using chi-square tests to assess associations between accuracy and study variables. Pairwise post hoc comparisons of column proportions were conducted using z-tests with Bonferroni correction.
Results:
A total of 27,300 responses were analyzed; 25,377 (92.96%) were correct. Question type significantly influenced chatbot performance (P<.001), with MCQs demonstrating the highest accuracy (5942/6300, 94.3%), and WH questions demonstrating the lowest accuracy (7686/8400, 91.5%). Language also showed a significant association with accuracy (P=.01), with English prompts achieving the highest accuracy (8512/9100, 93.5%). Chatbot model performance differed significantly (P<.001), with Gemini 3.1 demonstrating the highest accuracy (5176/5460, 94.8%; 95% CI 94.21%-95.39%), followed by ChatGPT-5.2 (5105/5460, 93.5%; 95% CI 92.85%-94.15%) and Kimi K2.5 (5071/5460, 92.9%). Prompt complexity and testing day did not significantly affect response accuracy (P=.69 and P=.96, respectively).
Conclusions:
Current LLM-based chatbots demonstrate high accuracy in answering structured questions related to dental medical emergency management. However, performance varies according to question format and language, and the presence of incorrect responses highlights potential safety concerns. AI chatbots may serve as educational tools, but their use as real-time decision support during dental medical emergencies should be approached with caution.