Before a child protective services worker arrives at your door, an algorithm may have already scored your family. It has reviewed your public benefits history, your prior involvement with the child welfare system, your housing instability, your substance abuse treatment records where accessible, and dozens of other data points drawn from government databases. It has produced a number between zero and twenty. That number—or a score like it, depending on the system your county uses—will influence whether the hotline call about your family gets screened in for investigation, and how urgently.
You will not be told the score exists. You will not be told what it is. You will not have the right to challenge it.
This is not a hypothetical future. It is the present operating reality in Allegheny County, Pennsylvania—where the Allegheny Family Screening Tool (AFST) has been in use since 2016—and in a growing number of jurisdictions across Arkansas, Colorado, Oregon, Tennessee, and elsewhere that have adopted their own versions of predictive risk modeling in child welfare.
What the Allegheny Tool Actually Measures
The Allegheny Family Screening Tool was developed by Erin Dalton of the Allegheny County Department of Human Services and researchers at the University of Chicago’s Chapin Hall. It is the most studied and most publicly documented predictive risk tool in child welfare in the United States, which is part of why it has become the central object of civil rights criticism in this space.
The AFST draws on more than 130 variables from county and state administrative data systems. Many of those variables are direct or indirect proxies for poverty: receipt of public housing assistance, participation in the Supplemental Nutrition Assistance Program (SNAP), Medicaid enrollment, prior involvement with the homeless services system. It also includes prior child welfare history—whether the family has previously been investigated, found indicated, or had children removed.
The tool’s developers have published validation studies showing it performs better than chance at predicting which cases will result in a subsequent abuse or neglect finding. Critics—including researchers at the AI Now Institute and the Data & Society Research Institute—have raised a fundamental objection: the tool may be predicting which families are surveilled more heavily, not which families are more dangerous. If Black and low-income families are already more likely to be reported to CPS, investigated, and found indicated—due to structural inequities in how poverty is reported and documented—then a model trained on that historical data will reproduce and amplify those inequities.
“The model learns from a dataset that reflects decades of racially biased intervention. When you train a prediction engine on biased outcomes, you get a biased prediction engine. That is not a bug. That is the math.”
The Race Problem in the Data
A 2020 analysis by researchers at the University of Pittsburgh found that Black children in Allegheny County were significantly more likely to be screened in for investigation than white children with similar risk scores on the AFST. The disparity was not explained by differences in family circumstances. It appeared to reflect the fact that the underlying data systems over-represent Black families due to higher rates of public benefits participation—itself a product of systemic economic inequality, not of parenting behavior.
Allegheny County acknowledged the findings and has made methodological adjustments to the tool over time. Critics argue that adjustments at the margins do not resolve the foundational problem: a predictive model that uses poverty as a proxy for danger will always over-flag poor families, and in a country where poverty is racially distributed the way it is in the United States, that means over-flagging Black, Indigenous, and Latino families at disproportionate rates.
The county has also resisted calls from the American Civil Liberties Union of Pennsylvania and from family advocacy organizations to allow families subject to AFST-informed decisions to see and challenge their scores. The ACLU-PA has argued that this opacity violates due process principles, particularly when the score influences a decision that can result in family separation—one of the most severe intrusions on liberty that government can impose without a criminal conviction.
The Expansion Beyond Allegheny
Allegheny County’s AFST was designed as a single-jurisdiction tool by researchers with academic accountability. The tools spreading across the country do not all meet that standard. Several are proprietary products sold by private vendors—including systems marketed under names like PredictAlign and Eckerd Connects’ Family Assessment tool—whose underlying models are trade secrets. Families, advocates, and even some courts cannot access the algorithmic logic that drives a risk score in those jurisdictions.
Arkansas deployed a predictive risk model developed by the Annie E. Casey Foundation and Evident Change (formerly the National Council on Crime and Delinquency). A 2023 audit by the Arkansas Division of Children and Family Services found that the tool was producing inconsistent results and that workers were uncertain about how to weigh its recommendations against their own professional judgment. The audit recommended additional training. It did not recommend suspension of the tool.
Oregon has used a structured decision-making tool in child welfare for more than a decade. Advocates in Portland have raised concerns that families in communities of color receive higher risk ratings at higher rates, and that caseworkers—who are under intense workload pressure and scrutiny following high-profile child death cases—are deferring to algorithmic scores rather than exercising independent judgment, precisely the opposite of what the tools are designed to support.
Due Process and the Right to Know
Federal law requires that state child welfare agencies operate with procedural safeguards under Title IV-E of the Social Security Act and the Child Abuse Prevention and Treatment Act (CAPTA). Neither statute explicitly addresses algorithmic decision-making tools. The Biden administration’s Executive Order on AI (October 2023) called for agencies to assess AI systems for bias and civil rights implications, but child welfare agencies at the state and county level are not directly bound by federal executive orders governing federal agencies.
In 2024, the Child Welfare Policy team at the Center for the Study of Social Policy published a set of model principles for ethical use of predictive analytics in child welfare, calling for transparency, explainability, and the right of families to know when algorithmic scoring has influenced a decision affecting them. As of this writing, no state has enacted legislation codifying those principles into law.
Several states—including California, New York, and Illinois—have enacted or proposed algorithmic accountability legislation in the broader context of government AI use. None of those frameworks has been applied specifically and comprehensively to child welfare risk scoring tools.
What Families Can Do
If you are subject to a CPS investigation and you believe an algorithmic tool may have influenced the decision to investigate your family, you have the right to request your case file under your state’s public records or child welfare access laws. That file may or may not include documentation of any risk score used. If it does not, your attorney can request it through discovery in any resulting legal proceeding.
Organizations doing active advocacy on this issue include the Family Defense Center (Illinois), the National Coalition for Child Protection Reform, and the ACLU chapters in Pennsylvania, Oregon, and Arkansas. If you have direct experience with algorithmic tools in child welfare proceedings—as a parent, a caseworker, an attorney, or a researcher—RFA wants to hear from you. Contact tips@radiofreeamerica.press. We protect sources and we read every message.
Note: RFA submitted public records requests to child welfare agencies in Allegheny County, Arkansas DCFS, and Oregon DHS regarding algorithmic tool documentation and racial impact assessments. Responses pending at time of publication.