Beyond Accuracy: Do Large Language Models Provide Clinically Adequate Responses to Parent-Initiated Questions on Pediatric Obsessive-Compulsive Disorder?

BackgroundParents increasingly use large language models (LLMs) to obtain information about child and adolescent mental health. In pediatric obsessive-compulsive disorder (OCD), where early recognition and intervention are critical, the accuracy of AI-generated information may have clinical implications. This study evaluated LLM responses to parent-oriented questions about pediatric OCD.MethodsAn exploratory cross-sectional evaluation was conducted using three conversational AI systems: ChatGPT (GPT-4o), Google Gemini, and DeepSeek. Twenty questions reflecting common parental concerns were identified through online sources and expert review and submitted with standardized prompts in separate chat sessions. Anonymized responses were independently rated by two child and adolescent psychiatrists for accuracy, reproducibility, and error patterns.ResultsChatGPT achieved the highest accuracy (≈91%) and reproducibility (≈89%), followed by Gemini (≈88%; ≈85%) and DeepSeek (≈86%; ≈83%). Domain-level analysis showed the highest scores for early recognition/diagnosis and treatment-related questions, and the lowest scores for questions on family functioning and daily life. Omission of clinically critical information-particularly insufficient emphasis on exposure and response prevention-was more common than factual inaccuracy.ConclusionsLLMs can provide accessible and generally accurate information for parents about pediatric OCD. However, variability across domains and limited prioritization of clinically essential content suggest they should complement, not replace, professional evaluation.