The Use of Artificial Intelligence across Child Protection Services: A Scoping Review

Background: Child maltreatment remains a persistent global public health challenge with profound consequences for children’s physical, cognitive, and psychosocial development. Child Protective Services (CPS) systems, traditionally reliant on professional discretion, are increasingly turning toward artificial intelligence (AI) to standardize outcomes and manage high caseload pressures. While the rapid advancement of machine learning and natural language processing offers potential for system enhancement, current applications in child welfare are often fragmented. This fragmentation complicates conclusions regarding how AI is utilized and its actual effects on service outcomes.

Objectives: This scoping review systematically maps the application of AI across the child maltreatment service continuum. It aims to synthesize current application areas and summarize the reported effects of AI-embedded services on CPS processes and outcomes.

Methods: Following PRISMA-ScR guidelines, eight electronic databases (including Scopus, Sociological Abstracts, PubMed, APA PsycINFO, Web of Science, Social Science and Social Work Abstracts, MEDLINE, and EMBASE) were searched for peer-reviewed studies published between 1991 and 2025. Eligible studies examined AI use in services designed to prevent or respond to child maltreatment for children under 18 years. Three reviewers with social work expertise independently screened 702 records. Ultimately, 27 studies met inclusion criteria and were synthesized using a narrative synthesis approach.

Results: Twenty-seven studies met the inclusion criteria. Most were conducted in high-income countries (85.1%), primarily the United States (n=14). While 15 studies involved child samples, most exceeding 1,000 participants, demographic reporting was notably low income; 13 studies failed to report any demographic characteristics, and only three child-focused studies reported racial composition. Regarding data sources, 48.1% of studies utilized clinical records, followed by textual data (18.5%) and image-based data (11.7%). Application areas were concentrated in early-stage decision-making: 17 studies focused on detecting or identifying maltreatment, while 6 focused on risk prediction. Methodologically, machine learning (ML) was most common (e.g., random forest, XGBoost), though recent studies have incorporated deep learning (DL) techniques like convolutional neural networks (CNNs) and retrieval-augmented generation (RAG). While most studies reported favorable performance with AI often exceeding the accuracy of human reviewers, one study found traditional statistical methods outperformed AI, and another intervention-stage study using AI-driven avatars reported non-significant improvements in professional training outcomes. Finally, ethical oversight was inconsistent, with 10 studies failing to report any legal or privacy-related information.

Conclusions and Implications: AI currently serves as a decision-support tool for early-stage CPS tasks, yet significant gaps remain. The review identified substantial methodological heterogeneity and a reliance on proxy outcomes (e.g., removal decisions) that may not accurately reflect actual child well-being. Furthermore, the frequent absence of demographic reporting on race and gender raises critical concerns about algorithmic equity and the potential reinforcement of systemic biases. For social work practice and policy, these findings suggest that while AI can augment professional capabilities, agencies should adopt equity-centered governance and "human-in-the-loop" methodologies. Future research should move beyond accuracy metrics to conduct implementation-focused evaluations that measure real-world impacts on service matching and long-term child safety.