Conversational AI for Child Abuse Detection Through Multistage Counseling: Model Development and Validation Study.

Background: Child abuse severely disrupts the healthy growth and development of children, resulting in long-term physical as well as emotional consequences. In real-world settings, despite the continuous increase in reported abuse cases, the chronic shortage of certified child abuse professionals has significantly increased the workload of individual counselors, making timely intervention increasingly difficult. Objective: This study aims to reduce the workload of counselors in real-world counseling settings by supporting counseling and child abuse detection processes. We propose a conversational artificial intelligence-based framework, Conversational Artificial Intelligence for Child Abuse Detection (CACAD), which conducts counseling with children and detects child abuse during the counseling process. Methods: CACAD uses a large language model (LLM) to conduct counseling and to detect 4 types of child abuse: neglect, emotional, physical, and sexual. During the question generation process, the LLM serves as the primary agent, supported by 2 auxiliary modules. In the process of counseling, the LLM first determines whether a child's response provides sufficient information to understand the current situation. If the response is deemed insufficient, the LLM generates follow-up questions to elicit additional information while preserving the context of the previous question. Once sufficient information is obtained, the next question category prediction module predicts the most appropriate category for the subsequent question and passes it to the LLM, enabling more flexible guidance of the counseling flow. In parallel, the abusive question detection module filters out potentially harmful or inappropriate questions to protect children. For abuse detection, CACAD uses an instruction-tuned LLM specialized for child abuse detection, and uncertainty quantification is applied to dynamically flag cases as pending review for counselor confirmation. Results: Experimental results using a Korean child-and-adolescent counseling dataset show that CACAD achieves strong performance in child abuse detection, with an exact match of 0.907 and a macro-F1-score of 0.939. In addition, CACAD demonstrates effective performance in counseling-related tasks, including next question category prediction and abusive question detection, contributing to coherent and safe counseling interactions. Human evaluation by domain experts further confirms the reliability of CACAD in counseling sessions, and uncertainty-based selective prediction allows the system to dynamically identify cases requiring human review. Conclusions: These findings demonstrate that LLM-based conversational agents can reliably perform child abuse detection within counseling conversations in real-world settings. The results further indicate that such systems can support counseling processes by integrating abuse detection, safe question generation, and uncertainty-aware decision handling in a unified framework.