When Prediction Becomes Policy - What Do Risk Models Actually Learn?

Explore how predictive risk models in child welfare systems function not merely as forecasting tools but as policy instruments that fundamentally reshape administrative practices and data generation. Examine the critical gap between theoretical model assumptions and real-world implementation through a comprehensive research program spanning systematic reviews, ethnographic studies, and computational analyses of child welfare agencies in Wisconsin and Ontario. Discover how risk predictors often encode agency responses and surveillance intensity rather than true risk indicators, and learn why the semantics of "risk" evolve throughout case lifecycles in ways that standard prediction frameworks fail to capture. Investigate the extension of these findings to contemporary LLM-based workflows, including research on service plan goal identification that reveals decreasing reliability as cases become more complex and documentation shifts toward emergent concerns outside formal service plans. Understand the implications for evaluation methodology by reframing assessment as an intervention problem, and gain practical insights for audit and design approaches that align predictive targets with accountable service pathways rather than standalone risk scores.