Smarter Safety: Leveraging AI to Strengthen Child Welfare Systems

Author: Christopher Mills and Susannah Petitt

A wall displays many vibrant children's drawings with rainbows, suns, moons, jellyfish, abstract shapes, and landscapes.

Download the Policy Brief

More evidence on predictive risk

Artificial intelligence tools have immense potential to improve child protective services, but their use is often framed as a choice between trusting an algorithm or trusting a human caseworker. This is a false dichotomy: the strongest results do not come from choosing one over the other, but from designing systems where algorithms and caseworkers are complements to improve outcomes for children and families. For example, a clear use case for algorithms is in summarizing information that agencies already collect – namely prior family child protection history – where workers have limited time to comb through details and where AI can refine assessments of risk or spot risk factors that humans can act on.

Better-targeted CPS investigations can meaningfully improve outcomes for families in crisis. When caseworkers have access to predictive analytics tools, they reallocate their attention toward children at greater risk of harm, without necessarily expanding the overall footprint of CPS involvement.

Designing Algorithms that Strengthen Families

Predictive analytics tools are already in use in child welfare agencies across the country, including in Allegheny County, Pennsylvania, Douglas County, Colorado, among others. Agencies should consider AI decision support when three conditions hold:

  • Human discretion remains central: Predictive analytics tools should be voluntary and advisory, or at least require and promote human engagement and discretion. AI should complement, not substitute for human expertise.
  • Implementation is transparent and evidence-driven: Workers should know what data is being used to feed the algorithm, what it predicted, and its limits. Workers should also be trained with concrete examples of when a score or other AI-flagged feature should not drive a decision.
  • Governance and culture fit: In one Colorado agency, larger decision-making groups saw greater equity gains from these tools. In many municipalities, screening and investigation decisions are made by individual workers who vary in experience and training. Agencies should consider how to best incorporate AI into their existing workflows, or to develop team structures that align the respective advantages of human workers and AI.

While algorithms can be extremely powerful tools, especially in complex, high-uncertainty referrals, they are only legally viable and most helpful if paired alongside human judgment and expertise. Child wellbeing, equity, and allocative efficiency may all be improved with the use of algorithms, but evidence points to them being most effective when overseen by workers who are trained in what the tool measures – and what it does not.

A common concern is that algorithmic tools will worsen racial and economic disparities in an already scrutinized system. Research instead shows that pairing caseworkers with algorithmic decision support reduced disparities in investigation rates compared to caseworkers acting alone. Algorithms are not guaranteed to advance equity, but human oversight of a well-designed tool can actively correct for bias in ways that neither humans nor algorithms necessarily achieve independently.

Agencies Should Experiment and Evaluate

Predictive risk tools can provide both improved outcomes for families, as well as a favorable return on public investment relative to implementation cost. Outcomes will vary based upon geography, demography, and staffing. In order to reap the benefits of predictive analytics, agencies should be rigorously evaluating their tools. Agencies should actively explore ways to organize workflows that maximize the complementary expertise of humans (collecting information, interpersonal connection with reporters and families, exercising oversight over final decisions) and AI (summarizing information and risk, helping ensure consistent standards, pointing out gaps in decision-making that helps to train workers). We welcome the chance to support jurisdictions and practitioners exploring predictive analytics in child welfare.