A Systematic Review of Artificial Intelligence Research within Child Welfare: Research-Driven Implications for Enhancing Practice

Background: Artificial intelligence (AI), including tools such as machine learning-based predictive analytics and natural language processing, has begun to show promise in the U.S. child welfare system. AI intends to simulate human decision-making through trained models, which can be used to generate or analyze text, images, or video. There have been a number of benefits of AI, including automation, improvements in decision-making or insight, reduction of risk or error, and increased access to services or supports. Despite these benefits, the integration of AI into child welfare practice also presents significant challenges, including concerns about algorithmic bias, lack of transparency, data quality issues, and limited stakeholder trust. The purpose of this research was to conduct a systematic review to understand methodological innovations, applications, and gaps in existing child welfare research.

Methods: The development of this systematic review was guided by the steps laid out in the PRISMA 2020 Checklist (Page et al., 2021). Authors utilized three databases covering a breadth of articles related to multiple forms of AI and child welfare/foster care. There were two independent coders at each stage of title/abstract screening and full-text review. Further details will be provided during the presentation related to: eligibility criteria, information sources and search strategy, selection and data collection process, data items, and synthesis methods.

Results: Following import, a total of 9,920 unduplicated articles were included for title/abstract screening, and 67 articles were narrowed down for full text review. Following full-text review, 57 articles remained. Articles reflected the following themes: AI to support professional decision-making; AI for detection, documentation, and information use; human judgment and trust in AI; ethics, bias, fairness, and rights; workforce readiness and organizational support; training and AI literacy; AI for family-centered practice and engagement; and data quality, evidence, and oversight. Most of the articles were focused on predicting risk of maltreatment or recurrence of maltreatment to inform decision-making. Specifically, decisions related to reporting maltreatment – including maltreatment prediction in medical systems, understanding factors leading to or exacerbating maltreatment, and identifying factors associated with service use or permanency. Specific findings will be discussed in more detail.

Conclusions/Implications: A number of gaps exist in the use, implementation, effectiveness, and ethical understanding of AI as a tool for supporting the child welfare workforce or families. There is a need for wider application and evaluation of AI tools across all topical or practice areas within child welfare, as well as long-term evaluation of the effects of AI use on families. Involving lived expertise and frontline workers in evaluation planning and analysis could also produce better ethical and practical understanding of AI tools. Additional implications and recommendations will be discussed related to supporting child welfare professionals, children families impacted by the system, and allied service professions.