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

Area of Science:

• Artificial Intelligence
• Child Psychology
• Clinical Informatics

Background:

• Child abuse has severe, long-term physical and emotional consequences.
• Increased reported cases and a shortage of professionals strain child abuse services.
• Timely intervention is crucial but challenging due to high workloads.

Purpose of the Study:

• To reduce counselor workload in child abuse detection and counseling.
• To introduce a conversational AI framework (CACAD) for child abuse detection.
• To support both counseling and abuse identification processes.

Main Methods:

• Utilizing a large language model (LLM) for counseling and detecting four types of child abuse (neglect, emotional, physical, sexual).
• Employing auxiliary modules for question generation, next question category prediction, and abusive question filtering.
• Implementing an instruction-tuned LLM with uncertainty quantification for reliable abuse detection and flagging cases for review.

Main Results:

• CACAD achieved high performance in child abuse detection (exact match 0.907, macro-F1 0.939) on a Korean dataset.
• Demonstrated effectiveness in counseling tasks like next question prediction and abusive question detection.
• Expert evaluation confirmed CACAD's reliability, with uncertainty-based prediction identifying cases needing human review.

Conclusions:

• LLM-based conversational agents can reliably detect child abuse during counseling.
• CACAD integrates abuse detection, safe question generation, and uncertainty handling into a unified framework.
• Such AI systems can significantly support counseling processes and improve child welfare outcomes.