LO
DeploymentCognitive automation2024-04-01
Allegheny County's own account of its child-welfare screening tool: supervisors override its default recommendation regularly, and the score is hidden once a case is opened
Social workeroccupation page →
Event date / reported
2024-04-01 · reported 2024-05-31
Evidence stage
DeploymentAn employer has put it into production. Can move the baseline — weighted by scale and how similar the setting is.
Tasks this bears on
Assessing risk and responding to crises
Judging whether a child or vulnerable adult is safe, visiting homes, and acting — including with statutory powers — when they are not.
Still human-led✓ Evidence-backed
Where this applies
Allegheny County's own Department of Human Services published this review on its own analytics site, synthesising outside peer-reviewed and working-paper research about the Allegheny Family Screening Tool (AFST), which the county has used since 2016 to help call screening caseworkers decide whether a report of suspected child abuse or neglect is investigated. The document is dated April 2024 and was posted 31 May 2024 — a first-party account of how the tool is actually used, not a vendor's description of what it can do. Three statements in it carry the judgement for this task. First, on the decision itself: the county writes that AFST risk information is used "alongside other information in the referral process," that intake supervisors "have discretion to override the defaulted selections and do so regularly," and that supervisors historically overrode about a quarter of mandatory screen-in recommendations. Second, on what happens after screening: once a referral is screened in for investigation, "neither the assigned investigator nor any caseworker ever sees the AFST score, so it does not affect ultimate assessments of maltreatment or decisions to open a case or remove a child from their home" — the score never reaches the statutory removal decision at all. Third, citing De-Arteaga et al. (2020): during a period when a software defect caused some scores to display incorrectly, screeners' actual decisions tracked the true underlying risk more closely than the erroneous displayed scores, which the county reads as evidence that staff apply independent judgement rather than following the displayed number. What this does not establish: caseload, headcount, how long a screening decision takes, or how any of this generalises beyond one county's child-welfare hotline. Allegheny County built, funds and operates the AFST itself, so a claim about its own restraint in using it is not the same as an outside audit reaching the same conclusion — though the cited academic studies are independent of the county.
What this means
The tool that flags risk at intake is not the same as the person who decides. Allegheny County's own account says a supervisor can and regularly does override the algorithm's default recommendation, and that once a case is opened for investigation, the risk score disappears entirely — the person who investigates, and any decision to remove a child, never sees it.
What it does not yet show
This is one county's own child-welfare hotline, not a measurement of caseload, headcount, or how long a screening decision takes, and it says nothing about social work outside this one setting. The county built and runs the tool itself, so its account of its own restraint is not an outside audit reaching the same conclusion.
What you can check
Ask whoever runs risk triage in your own setting one question: after the tool flags a case, who is the last person to touch it before an irreversible action, and can that person's name be found on the record?
Does it change the assessment?
No. The impact index is never moved by a single event. What this record did: the 1 linked task judgement above now rest on evidence instead of inference.
Source
Allegheny County Department of Human Services — Summarizing Recent Research on Predictive Risk Models in Child Welfare (April 2024) · verified 2026-09-24 · VOLO agent loop · interpreted 2026-09-24 · VOLO agent loop
Primary source — published by the party that did this, or the authority of record. No co-signature needed.