RESEARCH OBJECTIVES: Social service risk assessments in Sweden have been criticized for lacking transparency. The present project aims to alleviate cognitive limitations in these assessments by creating an AI-based support system.

THEORETICAL BACKGROUND: Previous research has shown that human cognition is limited. Custody evaluators are expected to ensure that all components are included and understood, include relevant scientific and experience-based information, and merge all information into transparent risk assessments. A generative AI-system was developed using role-based prompt engineering and ChatGPT 4o. The AI-system contains a support structure and scientific information relevant to social service risk assessments. The AI-system, via chat-bot dialogue, guides the social worker through the aforementioned model and information. Upon completion, the AI-system suggests a preliminary risk assessment based on a mathematical model but leaves the final decision to the user.

METHOD: Iterative design was used to measure usability where two active custody evaluators tested the AI-system, met the developer, and verbally commented on the AI-system. The results are based on the second iteration.

RESULTS: Testers appreciated memory overload amelioration, an objective preliminary risk assessment, and increased transparency, time efficiency, and learning curve. Also, dialogue rigidity and oversimplification in the risk model were stressed. Testers desired support in removing irrelevant information, using the AI system as a litigation deterrent for parents, and adding relevant information.

LIMITATIONS: External validity is limited because of few testers and the system is only in its second iteration.

RESEARCH IMPLICATIONS: The AI-system provided structure and information important for increased transparency in social service risk assessments, and allowed faster learning for inexperienced custody evaluators. Scientific validation of transparency and overall quality is needed.