Algorithmic Bias in Predictive Models for Child Protective Services

Abstract

Predictive risk models such as the Allegheny Family Screening Tool are increasingly used to help child protective services screeners decide which referrals warrant investigation, on the premise that a statistical score can outperform inconsistent human judgment under severe time pressure. This policy analysis reviews the ethical and empirical literature on how these tools are built and validated,
the specific mechanisms by which they can encode racial, disability, and poverty-linked disparities, and what happens when their scores are handed to human decision-makers rather than acting alone. The evidence indicates that variables correlated with race, disability, and poverty function as proxies for risk rather than direct measures of it, that human override of algorithmic scores has itself been shown to introduce racial disparity rather than reliably correct it, and that most deployed tools have not been evaluated
against fairness-reporting standards that would make these problems visible before deployment. We examine governance proposals converging on independent pre- and post-deployment fairness auditing, deliberate scrutiny of proxy variables, genuine institutional support for caseworker discretion, and mechanisms for family transparency and contestability. We conclude that no currently deployed system satisfies all of these conditions, and that the field's own reviewers still describe the underlying validation problem, testing a model against outcomes that a historically biased system itself produced, as unresolved.