Predictive analytics in child welfare: Promise, peril, and the path forward

As the government shows major investments for predictive analytics in child welfare, the field now stands at a crossroads. The lessons of early adopters, cautionary legal tales, and emerging international models aid in charting a more responsible road ahead.

professional photo of Jamey smiling

Jamey Welty

Service Delivery Manager

banner image

A federal signal of change in the child welfare system: The ACYF Demonstration Grant

Child welfare agencies across the United States have long wrestled with a difficult truth: the decisions their workers make (often under intense time pressure and with incomplete information) carry life-altering consequences for children and families. Increasingly, policymakers and practitioners are asking whether technology can help close that gap.

The Administration for Children, Youth and Families (ACYF), a division of the federal Administration for Children and Families (ACF), is now clearly signaling which direction it wants the field to move. In 2026, ACF announced $6 million in competitive grant funding under the Predictive Analytics in Child Welfare Demonstration Grants program (Funding Opportunity Number HHS-2026-ACF-ACYF-CA-0037). This initiative invites child welfare program jurisdictions to serve as demonstration sites for the responsible design, implementation, and evaluation of predictive analytics tools, with findings intended to support national replication.1

The announcement did not emerge in a vacuum. ACF held a roundtable in December 2024 with child welfare directors, researchers, and advocates to discuss the potential and current state of predictive analytics. That conversation led to the publication of an issue brief in March 2025, “Modernizing Child Welfare Technologies and Tools: Opportunities for Predictive Risk Modeling to Improve Child Safety and Outcomes”, which laid the intellectual groundwork for the grant opportunity.2

Most child welfare agencies in the United States still depend on manual tools to assess children’s risk of abuse and neglect—tools that are time-consuming, prone to human error, and rely on predetermined question weights with minimal predictive accuracy.

ACF Issue Brief, March 20253

The demonstration allows direct grantees to accomplish several goals simultaneously:

  • Develop and implement predictive risk models tailored to local data and needs
  • Train staff in responsible use of analytics tools
  • Establish governance structures and quality assurance processes
  • Engage communities and stakeholders in implementation
  • Rigorously evaluate and share lessons learned

The framing of the grant reflects how much the federal posture has evolved. Rather than simply mandating or endorsing algorithmic tools, ACF is explicitly investing in responsible governance, community engagement, and outcome evaluation as core requirements, not optional add-ons. Individual awards are expected to range from $400,000 to $600,000, with a total program investment of $6 million.4 The agency appears to have absorbed the hard lessons that early adopters learned in the field, and it wants demonstration sites to build those lessons in from the start. For state and local child welfare systems that have long recognized the promise of data-driven decision support but lacked the resources, this is a promising opportunity.

Responsible child welfare system innovation in Allegheny County

No jurisdiction has contributed more to what the field knows about building predictive analytics tools than Allegheny County, Pennsylvania. Since 2016, the Allegheny County Department of Human Services deployed the Allegheny Family Screening Tool (AFST), a machine learning model that draws on integrated data from Medicaid records, behavioral health, criminal justice, and school systems to generate a risk score for each child referred to the child welfare hotline. What set Allegheny County apart was its insistence on building in accountability before deployment. The process included community meetings, a formal validation study, an independent impact evaluation (especially how this would impact existing child welfare policies), and an ethical review prior to launch. County officials published detailed documentation of the model’s methodology so independent researchers could scrutinize it; a level of transparency that was unprecedented in child welfare system technology at the time.5

The AFST’s design reflects a core principle that has since become standard in the field. The tool augments, rather than replaces, caseworker judgment. Screeners receive the risk score only after forming their own initial assessment, and supervisors retain authority to override algorithmic thresholds. Research found that the tool meaningfully shifted the composition of investigated referrals toward higher-risk cases.6 Evaluation found that the Black-White disparity in screen-in rates declined from 9 percent before the AFST to 7 percent after its introduction, with similar reductions observed in case opening and home removal rates for investigated referrals; reductions achieved, researchers noted, while improving outcomes for both Black and white children.7

Allegheny County did something that very few jurisdictions have been willing to do. They built the tool in public, invited independent scrutiny, published the results, and kept iterating. That willingness to be examined is exactly what the field needs to get this right.

Because the AFST was built in the open, it also drew rigorous critical research, generating insights the whole field has benefited from. Independent analyses raised important questions about training outcome selection. It was noted that a model designed to predict removal (a historical child welfare system response) rather than actual maltreatment can encode existing disparities into its predictions.8 Research also found that when screeners overrode algorithmic scores, those overrides often improved equity outcomes, confirming that practitioner judgment and algorithmic tools work best as genuine complements.9 A U.S. Department of Justice review further clarified that algorithmic tools in child welfare carry the same civil rights obligations as any government decision process; a clarification that now informs implementation standards across the country.10

The AFST also clarified important obligations at the intersection of predictive tools and disability rights. Families raised concerns that diagnoses fed into the algorithm through Medicaid and behavioral health records appeared to drive elevated scores even without evidence of maltreatment, prompting a U.S. Department of Justice review in 2023. Rather than a condemnation, the field can read that scrutiny as a necessary clarification. Civil rights frameworks apply fully to algorithmic tools, and proactive civil rights review must be a standard feature of implementation, not a response triggered after complaints.11

The AFST experience also showed the direct link between explainability and worker trust. When caseworkers received limited information about how scores were generated, many developed informal theories about the model’s logic, sometimes collaborating to reverse-engineer its behavior. That impulse is constructive. Practitioners want to engage with these child welfare system tools, not merely receive scores. Implementations that invest in plain-language explanations, and structured feedback channels, can channel that energy into a more effective partnership between human expertise and algorithmic analysis.12

Florida’s Eckerd Rapid Safety Feedback focuses on open cases

While Allegheny County focused its tool on the intake and screening function, Eckerd Kids, a Florida-based nonprofit, developed the Eckerd Rapid Safety Feedback (ERSF) process with a different aim. Their goal was to identify high-risk children among cases that are already open and under investigation. The ERSF combines a predictive risk model with a real-time quality assurance protocol, in which specialized reviewers provide coaching to frontline investigators handling flagged cases.13

The model identifies risk factors associated with child fatalities, including children under the age of three, the presence of a paramour in the home, domestic violence history, and substance abuse. Cases meeting certain risk thresholds trigger immediate supervisory review and an intensified safety planning process. The ERSF approach gained national attention and was highlighted in a 2016 report by the Commission to Eliminate Child Abuse and Neglect Fatalities This report called out Hillsborough County, Florida, as a model of data-driven child welfare protection practice.14

By 2016 and 2017, seven states were working toward implementation of ERSF. However, a rigorous multi-year evaluation published in 2022 in the journal Child Abuse and Neglect found no statistically significant reduction in subsequent severe maltreatment among children flagged by the algorithm. Researchers noted that future maltreatment within 12 months of prior investigation may be too distal an outcome for a process designed primarily to improve investigative practice quality.15

New Zealand’s deliberately cautious approach to predictive analytics in child welfare

New Zealand has the distinction of being among the first countries in the world to seriously explore predictive risk modeling and predictive analytics in child welfare. The country is also one of the most deliberate in its caution about moving from research to practice. More than a decade ago, researchers at Auckland University of Technology demonstrated that integrated administrative data could identify newborn children at statistically elevated risk of later maltreatment.16

The New Zealand government’s Ministry of Social Development commissioned multiple feasibility studies and privacy impact assessments, conducted an ethical review focused specifically on implications for the Māori community, and engaged Statistics New Zealand in a peer review process. That review concluded that a model should trigger closer professional assessment rather than automatic intervention, a principle that has guided New Zealand’s cautious approach ever since.17

The pathway from promising research to operational practice has nonetheless proved contentious in New Zealand. Critics have raised concerns about surveillance of vulnerable families, the risk of false positives leading to unnecessary intervention, and the disproportionate impact on indigenous communities (whose historical overrepresentation in administrative datasets could skew model predictions). Those concerns have kept New Zealand in a deliberate testing and simulation mode, even as child protection workers continue to face significant caseload pressures.18

What 3 jurisdictions taught the field

Allegheny County, Florida, and New Zealand approached predictive analytics in child welfare from different angles, yet their experiences surface a consistent set of lessons that should anchor every future implementation.

  • Transparency drives field learning. Allegheny County’s decision to publish its methodology, and invite independent evaluation, generated the research base the entire field is now drawing from. Jurisdictions that build in the open shorten the learning curve for everyone else.
  • Outcome selection is a foundational equity decision. The AFST’s reliance on historical removal as a proxy for maltreatment risk, and the ERSF’s use of a recurrence measure too distal from the intervention to detect effects, both show that what a model is trained to predict determines whether it advances or encodes inequity.
  • Evaluation turns implementation into knowledge. The ERSF found no significant effect on its target outcome; the AFST found equity gains alongside unresolved tensions. Neither finding constitutes a failure because both are exactly what the field needs. Tools in a child welfare system that are not evaluated cannot improve.
  • *Human judgment and algorithmic tools thrive in partnership. From Allegheny County’s screener override patterns to New Zealand’s principle that model output should trigger professional assessment rather than automatic action, every implementation reinforced by the practitioner remains the decision-maker. Tools in a child welfare system that are designed to inform that judgment, rather than replace it, produce better outcomes and can sustain greater worker and community trust.
  • Governance must precede deployment. New Zealand’s methodical ethical review, Allegheny County’s pre-launch community process, and the DOJ’s civil rights examination of the AFST all point to the same conclusion. The time to resolve questions of equity, explainability, and civil rights compliance is before a tool goes live, not after families and workers raise alarms.

These lessons do not argue against using predictive analytics in child welfare; they argue for using them with discipline, transparency, and genuine accountability to the communities most affected. That is precisely the standard the ACF demonstration grants are designed to establish. Now, the question is how to translate these hard-won insights into a new generation of child welfare policies. These implementations should expand on what succeeded, avoid what did not, and expand use cases the field has barely begun to explore.

How predictive analytics in child welfare can be used responsibly

The following principles and use cases represent promising directions for responsible application of predictive analytics in child welfare, informed directly by the lessons the field has accumulated.

  1. Support, not replace, human judgment

    Predictive tools work best, and cause the least harm, when they are positioned as one input among several in a structured decision-making process. Tools that mandate specific outcomes based on algorithmic scores, or that present risk scores without adequate context, undermine practitioner agency. Future implementations should embed predictive analytics in child welfare within structured decision frameworks, requiring workers to document their reasoning when deviating from a model’s recommendation in either direction.

  2. Move from static tools toward dynamic, learning models

    Most child welfare agencies currently rely on the Structured Decision Making (SDM) model, an actuarial framework that uses fixed, predetermined weights to classify families by risk level at key case decision points. SDM represented a genuine advance over unaided clinical judgment research consistently found that structured actuarial tools outperform unassisted worker assessments in predicting future maltreatment. However, SDM and similar static tools carry important limitations that the field is increasingly positioned to address.19

    Static tools assign the same weights to risk factors, regardless of local context, population shifts, or changes in agency practice over time. They cannot update their predictions as new information enters a case, and they do not learn from outcomes.20 Research also found that risk, as it appears in caseworker notes and practice, is a temporally dynamic construct. It changes as family circumstances change, as services are engaged, and as protective factors strengthen or erode. A fixed checklist scored at intake cannot capture that movement.21

    Dynamic machine learning models, by contrast, can be:

    • Retrained on current local data
    • Updated as populations and child welfare policies evolve
    • Designed to incorporate real-time information about a family’s trajectory, rather than a single point-in-time snapshot

    The ACF demonstration grants represent an opportunity to develop and test models that complement the structured decision-making framework rather than simply replace it. Child welfare system tools that retain SDM’s strengths of consistency, accountability, and clear decision points add the predictive power, adaptability, and data depth that static actuarial tools cannot provide. When this is done correctly, the next generation of tools can provide workers with a more detailed, current picture of family risk and protective factors at every stage of a case.22

  3. Shift the focus toward prevention and family support

    The field has invested heavily in predictive models designed to identify when to investigate or remove a child. There is a compelling, and comparatively underexplored, opportunity to use the same data infrastructure to identify families who would benefit from voluntary, preventive services before a crisis occurs. Predictive tools that flag families as eligible for home visiting programs, parent support services, or substance abuse or housing resources represent a fundamentally different use case. This methodology is oriented towards family preservation, rather than surveillance and enforcement. This aligns with the broader federal push toward prevention-oriented child welfare policies under the Family First Prevention Services Act, and it substantially reduces the civil rights risks associated with investigation-focused tools.23

  4. Build in bias auditing and algorithmic accountability from day 1

    Any jurisdiction pursuing predictive analytics in child welfare under the ACF demonstration grants, or outside them, should treat equity auditing as a core design requirement. This means selecting child welfare system training outcomes that reflect actual child safety, rather than proxy measures that encode historical disparities (disaggregating model performance by race, disability status, income, and geography at regular intervals). This also means establishing an independent review body with authority to require model modifications or suspend deployment if disparate impact is identified. Transparent documentation of model inputs, weights, and performance metrics should be publicly accessible, enabling community oversight and academic scrutiny.

  5. Invest in workforce preparation and explainability

    Predictive tools will only improve outcomes if the practitioners using them understand what the tools are measuring, what they are not measuring, and how to integrate scores with their own professional observations. Workforce development must be treated as an ongoing investment, not a one-time training during implementation. Agencies should work with vendors and researchers to develop plain-language explanations of model logic that can be shared with both workers and families. In jurisdictions where parents are legally entitled to know what information is being used to make decisions about their children, algorithmic inputs should also be disclosed proactively.24

  6. Use analytics to strengthen foster care placement and reunification

    Beyond initial investigation and screening, predictive analytics in child welfare holds genuine promise in optimizing placements. It can match children with foster families whose characteristics align with a child’s specific needs and predict which reunification plans are most likely to result in stable, safe outcomes. ACF’s own issue brief identifies caregiver recruitment and matching as a priority area for analytic investment. These child welfare system applications are less fraught than investigation-focused tools, because they are oriented toward improving service quality rather than triggering enforcement action, and they carry a different risk profile with respect to civil rights exposure.25

  7. Establish cross-jurisdictional learning networks

    The ACF demonstration grants explicitly require grantees to participate in collaboration and shared learning activities. This is one of the most valuable design features of the program. The child welfare system in the US is a decentralized system with no single national model and different structures by state for state and county functions, which means that lessons learned in one jurisdiction rarely transfer efficiently to others. Structured learning networks, supported by technical assistance from ACF and research partners, can change that. This network collaboration can accelerate the identification of what works, surface common failure modes, and help jurisdictions avoid reinventing problematic approaches that have already caused harm elsewhere.26

  8. Center community voice and family engagement

    Perhaps the most consistent failure mode across all early implementations of predictive analytics in child welfare has been the absence of meaningful community engagement. Families who are subject to these tools, disproportionately low-income families and families of color, have rarely had a seat at the table when models were designed, validated, or governed. The demonstration grants require community and stakeholder engagement as a grantee activity. Child welfare agencies should go further, establishing formal community advisory structures with genuine authority to shape model design, review performance data, and trigger independent audits. When the communities most affected by these tools have a voice in how they are built and governed, the likelihood of equitable outcomes increases substantially.27

A demonstrative moment in the child welfare system

The ACF demonstration grants have arrived at a crucial moment. The field now has nearly a decade of operational experience with predictive analytics in child welfare, enough experience to identify the conditions under which these tools create value and the conditions under which they cause harm. The question is no longer whether to use data-driven decision support in child welfare, but how to do so in a manner that genuinely improves safety, preserves family integrity, and advances equity rather than undermining it.

The jurisdictions that receive demonstration funding will have the opportunity, and the responsibility, to build the child welfare system models that the rest of the country will eventually follow. If this is done the right way, this generation of implementations can establish governance frameworks, child welfare policies, equity standards, and workforce practices that make predictive analytics a genuine asset for vulnerable children and families.

The federal investment is a vote of confidence that the child welfare field is ready to get this right. There’s ample opportunity to learn from these lessons; the question is whether the field will use them.

To learn about how CAI partners with health and human services organizations, fill out the form below.

Endnotes

  1. Administration for Children and Families. “Predictive Analytics in Child Welfare Demonstration Grants (Funding Opportunity No. HHS-2026-ACF-ACYF-CA-0037)”. U.S. Department of Health and Human Services. 2016. https://simpler.grants.gov/opportunity/0990b2d9-ef71-4e1f-9f97-6dfe3a9743a0. ↩
  2. Administration for Children and Families, Office on Child Abuse and Neglect. “Modernizing Child Welfare Technologies and Tools: Opportunities for Predictive Risk Modeling to Improve Child Safety and Outcomes.” U.S. Department of Health and Human Services. March 2025. https://acf.gov/acyf/policy-guidance/modernizing-child-welfare-technology-predictive-risk-modeling. ↩
  3. Administration for Children and Families, Office on Child Abuse and Neglect. “Modernizing Child Welfare Technologies…” U.S. https://acf.gov/acyf/policy-guidance/modernizing-child-welfare-technology-predictive-risk-modeling. ↩
  4. Administration for Children and Families. “ACF Announces $6 Million for States to Pilot Predictive Analytics in Child Welfare.” U.S. Department of Health and Human Services. May 2026. https://acf.gov/media/press/2026/acf-announces-6-million-states-pilot-predictive-analytics-child-welfare. ↩
  5. Allegheny County Department of Human Services. “Evaluation Findings on the Use of Predictive Risk Models in Child Welfare.” Allegheny Analytics. 2024. https://analytics.alleghenycounty.us/2024/05/31/predictive-risk-models-in-child-welfare/. ↩
  6. American Bar Association, Children’s Rights Litigation Committee. “Algorithmic Decision-Making in Child Welfare Cases and Its Legal and Ethical Challenges.” ABA Litigation Resources. Winter 2024. https://www.americanbar.org/groups/litigation/resources/newsletters/childrens-rights/winter2024-algorithmic-decision-making-in-child-welfare-cases/. ↩
  7. Rittenhouse, K., Goldhaber-Fiebert, J. D., & Prince, L. “Algorithms, Humans and Racial Disparities in Child Protection.” 2023. https://krittenh.github.io/katherine-rittenhouse.com/AFST_Disparities.pdf. ↩
  8. Ballantyne, N. “The Harm That Data Do: The Case of the Allegheny Family Screening Tool.” Medium. August 2023. https://medium.com/@neilballantyne/the-harm-that-data-do-the-case-of-the-allegheny-family-screening-tool-5f9fca22e0b2. ↩
  9. Samant, A., Shemtov, N., Xu, K., Beiers, S., Gerchick, M., Gutierrez, A., Horowitz, A., Jegede, T., & Shah, T. “The Devil is in the Details: Interrogating Values Embedded in the Allegheny Family Screening Tool.” Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, 1292–1310. Human Rights Data Analysis Group. 2023. https://hrdag.org/2023/06/22/afst/. ↩
  10. Ho, S., & Burke, G. “Child Welfare Algorithm Used by Allegheny County DHS Faces Justice Department Scrutiny.” Associated Press / WESA. January 31, 2023. https://www.wesa.fm/politics-government/2023-01-31/child-welfare-algorithm-used-by-allegheny-county-dhs-faces-justice-department-scrutiny. ↩
  11. Sullum, J. “Child Welfare Algorithm May Unfairly Target Disabled Parents, Complaints to DOJ Allege.” Reason. February 2, 2023. https://reason.com/2023/02/02/child-welfare-algorithm-may-unfairly-target-disabled-parents-complaints-to-doj-allege/. ↩
  12. Kawakami, A., Stapleton, L., Lee, M. H., Qing, D., Wright, M., Chouldechova, A., Holstein, K., Wu, Z. S., & Zhu, H. “Algorithmic Harms in Child Welfare: Uncertainties in Practice, Organization, and Street-Level Decision-Making.” 2022. https://arxiv.org/pdf/2308.05224. ↩
  13. Oklahoma Department of Human Services. “DHS Partners with Tom Ward and Eckerd Kids to Bring New Technology to Child Protective Investigations”. June 23, 2016. https://oklahoma.gov/okdhs/newsroom/2016/june/comm06232016.html. ↩
  14. Eckerd Connects. “Eckerd Rapid Safety Feedback Highlighted in National Report of Commission to Eliminate Child Abuse and Neglect Fatalities.” 2016. https://eckerd.org/eckerd-rapid-safety-feedback-highlighted-national-report-commission-eliminate-child-abuse-neglect-fatalities/. ↩
  15. Lorthridge, J., McCroskey, J., Pecora, P. J., Chambers, R., & Fatemi, M. “Examining the Effects of the Eckerd Rapid Safety Feedback Process on the Occurrence of Repeat Maltreatment Among Children Involved in the Child Welfare System.” Child Abuse and Neglect, 133, Article 105858. 2022. https://doi.org/10.1016/j.chiabu.2022.105858. ↩
  16. Mirage News. “New Zealand Child Protection Strain: Can Predictive Tech Aid?” March 22, 2026. https://www.miragenews.com/new-zealand-child-protection-strain-can-1641810/. ↩
  17. New Zealand Ministry of Social Development. “Vulnerable Children Predictive Modelling.” https://www.msd.govt.nz/about-msd-and-our-work/publications-resources/research/predicitve-modelling/. ↩
  18. Barmomanesh, S., & Miranda-Soberanis, V. “Not on My Watch! A Case Study in the Datafication of Child Welfare in Aotearoa New Zealand.” ResearchGate / University of Auckland. 2023. https://www.researchgate.net/publication/356149454. ↩
  19. Saxena, D., Badillo-Urquiola, K., Wisniewski, P. J., & Guha, S. A Human-Centered Review of Algorithms Used within the U.S. Child Welfare System. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. https://arxiv.org/pdf/2003.03541. ↩
  20. Bartelink, C., van Yperen, T. A., & ten Berge, I. J. “Determining Child Maltreatment: A Meta-Analysis on the Predictive Validity of Risk Assessment Instruments.” Child Abuse and Neglect, 99, Article 104171. 2020. https://doi.org/10.1016/j.chiabu.2019.104171. ↩
  21. Saxena, D., Kawakami, A., Rolnick, J., & Chouldechova, A. “Beyond Predictive Algorithms in Child Welfare.” 2024. https://arxiv.org/pdf/2403.05573. ↩
  22. Saxena, D., et al. “A Human-Centered Review of Algorithms…” https://arxiv.org/pdf/2003.03541. ↩
  23. Administration for Children and Families, Office on Child Abuse and Neglect. “Modernizing Child Welfare Technologies and Tools...” https://acf.gov/acyf/policy-guidance/modernizing-child-welfare-technology-predictive-risk-modeling. ↩
  24. Kawakami, A., et al. “Algorithmic Harms in Child Welfare…” https://arxiv.org/pdf/2308.05224. ↩
  25. Administration for Children and Families, Office on Child Abuse and Neglect. “Modernizing Child Welfare Technologies and Tools...” https://acf.gov/acyf/policy-guidance/modernizing-child-welfare-technology-predictive-risk-modeling. ↩
  26. Administration for Children and Families. “Predictive Analytics in Child Welfare…” https://simpler.grants.gov/opportunity/0990b2d9-ef71-4e1f-9f97-6dfe3a9743a0. ↩
  27. Samant, A., et al. “The Devil is in the Details…” https://hrdag.org/2023/06/22/afst. ↩

Related resources

The transformational power of ML and AI in government
Thought Leadership

The transformational power of ML and AI in government

Cutting edge technologies are changing the way public agencies operate.

View this resource →
Enhancing child welfare: Best practices for measuring and improving outcomes
Webinar On-Demand

Enhancing child welfare: Best practices for measuring and improving outcomes

View this on-demand webinar focusing on child welfare specifically designed for county governments, with lessons that can be applied to other behavioral health programs.

View this resource →
Collaborative governance for delivery of successful human services
Webinar On-Demand

Collaborative governance for delivery of successful human services

Watch a 45-minute webinar on collaborative governance, which connects siloed stakeholders for coordinated and efficient service delivery and improved outcomes.

View this resource →

Let's talk!

Interested in learning more? We'd love to connect and discuss the impact CAI could have on your organization.

All fields marked with * are required.
Please correct all errors below.
Please agree to our terms and conditions to continue.

For information about our collection and use of your personal information, our privacy and security practices and your data protection rights, please see our privacy policy and corresponding cookie policy.