Child welfare & family services

The most-studied predictive risk score in child welfare — and evidence that the human override layer is where equity was won or lost.

A child-welfare risk tool that flagged thousands of children at extreme risk while missing actual fatalities — discontinued in 2017.

An Allegheny Family Screening Tool (AFST)-derived screening tool that Oregon shelved in 2022 amid equity concerns — a rare pre-crisis discontinuation.

A council's vendor-built family-profiling pilot, run without telling the families, quietly scrapped in 2019 after data-quality problems meant it surfaced few genuinely new cases.

A predictive model scores every newborn in the county to offer prevention services, not investigations — benevolent intent riding on a maltreatment-removal prediction, delivered by mandated reporters behind a deliberate firewall.

An independently-trialed, race-excluded child-welfare screening score that a rigorous evaluation found sped decisions but barely changed outcomes — because workers heeded only the extremes.

Eckerd Rapid Safety Feedback, a child-welfare risk tool whose Florida success was the vendor's own claim — endorsed by a federal commission and copied across states years before an independent evaluation found it changed nothing.

Dutch police risk-profiling that scored children under 12 — including those recorded only as victims — from up to twelve years of police records; the government's own pilot evaluation found no well-functioning instrument in any of its four regions.

A population-scale police-and-council database whose child-exploitation risk models were quietly switched off as 'not fit for operational use' — leaving auditors unable to find the code, the variables, or any record of why.

A national predictive-risk score that ranks children for early-help outreach from data families gave to receive benefits — its operational accuracy never published, its independent bias audit never disclosed.

A proprietary child-abuse risk model that a 95.6% false-positive (people flagged who turned out not to be cases) rate stopped in retrospective testing — shelved before it ever scored a live case.

A government-funded evidence centre built child-welfare prediction models, pre-registered a success bar, missed it on every model, and published the failure — the rare case where the evidence gate fired before deployment.

A national child-maltreatment risk model a minister stopped before it ever ran — among the Atlas's never-deployed tools, the one halted by an accountable authority rather than by a test result.

A whole-population child-vulnerability model halted in development and never run — undone not by its accuracy but by legal basis, legitimacy, and an unrelated data breach.

Since May 2018, New York City's child-welfare agency has scored every open child-protection investigation at day 10 with an in-house machine-learning model, rank-ordering cases by predicted likelihood of substantiated physical or sexual abuse within 24 months to fill a capacity-limited 'quality assurance' review worklist — about 3,000 of roughly 50,000 investigations a year. Nobody who acts on the output sees it: not families, not their attorneys, not frontline caseworkers, and — by the city register's own words — not even the QA reviewers working the flagged list. The agency's own internal audit conceded the training data likely carried systemic biases, that geography may act as a partial race proxy, and that flag predictions are 'more likely to be incorrect than correct'; the agency told state auditors the model outperforms experienced caseworkers (a self-claim with no public methodology) and that there is 'no basis for a complaint.' The oversight graph is blocked-but-not-inert: a 2023 state audit produced promises, the charter-mandated inspector general is statutorily severed from the records — full access in only 1 of 18 child fatalities with prior agency involvement in 2025 — and the algorithmic-accountability office the Council legislated in November 2025 remains unverified as operational, while the model expanded in 2025 into the very record class state law prohibits the inspector general from touching.

Colorado runs every screened-in child welfare referral through two statewide instruments: a structured Family Safety Assessment (10 danger items gated by 5 judgment-heavy threshold criteria) and an actuarial Family Risk Assessment scored largely from prior system contact stored in the Trails record system, used by 64 county-administered agencies under one state rule. In 2024 the legislature handed the state's independent Child Protection Ombudsman — not the child welfare agency — a funded mandate (HB 24-1046, $109,392) to procure a third-party audit of the tool suite against nine statutory criteria. ICF's 176-page audit (released March 2, 2026; 50 recommendations) found the tools well-aligned with policy on paper but inconsistently implemented: 68% of caseworkers cannot complete assessments in real time, vignette agreement collapses to 59–63% on neglect and poverty scenarios, the risk tool's weighting of old reports 'disproportionately impacts minority families and perpetuates systemic bias' (with any past DV incident coded as recent), and race and ethnicity are documented so inconsistently that the disparity analysis the statute requested could not be run. The audit recommends revising or replacing the actuarial tool — and as of July 2026 the oversight loop has produced statute, budget, and public information, but no tool change.

Since 2018, Adoption-Share's Family-Match — a proprietary 'relational fit' matching algorithm built by former eharmony researchers — has paired foster children awaiting adoption with self-recruited prospective parents in Florida, Georgia, and Virginia, emitting a ranked candidate-parent list per child that caseworkers then vet toward trial placement and legal adoption. The independently observed results are stark: Virginia's two-year test produced one known adoption (an official's statement to the AP; the pilot's end is undated in the record, circa 2020) and Georgia's year-long pilot produced two before ending in October 2022 — yet both states resumed business with the vendor, Georgia for free in July 2023 after direct lobbying, Virginia via a larger contract for a rebranded recruitment portal ($212,546 spent in SFY2023). In Florida, the showcase state, the vendor's own ledger claimed 603 placements and 431 adoptions over five years — figures its partner agencies could not verify: one agency's records showed 76 tool-attributed placements with no documented adoption, another counted 8 adoptions from 22 matches while making hundreds of both without the tool. The AP's November 2023 investigation (Sally Ho and Garance Burke) found the vendor owned the outcome data, refused states access to the algorithm even on request, and — per its own April 2023 user guide — had caseworkers document matches made outside the tool inside the system. Only Tennessee interrogated the data schema before deployment; only Tennessee never deployed.

Five US states run the child-welfare domain's only store-to-store trigger: a record join between new birth certificates and a permanent registry of parents with prior terminations, with no risk score and no screening discretion — a self-arming automation so unowned that match volumes halved in two states whose agencies were unaware of the collapse or unable to explain it, and the one legally mandated evaluation was never performed.

In June 2025, DC's Child and Family Services Agency shipped the child-welfare domain's first documented policy-retriever deployment: CORA, a staff-facing chatbot embedded in STAAND, the agency's new case-management system, answering policy and procedure questions from a small expert-validated corpus — with a published 16-page pre-deployment governance report (one of only three district-wide) whose centerpiece guardrail is structural: the chatbot 'does not draw from confidential case files' and has no access to STAAND case data. Within six months the agency's own tip sheets documented the guardrail's blind side: a CORA phone app that ingests case documents, photos of handwritten notes, and voice dictation, and saves AI-drafted contact notes into the STAAND case record — the write direction the no-case-files rule never governed, superseding the May 2025 report's claim that generated text was 'guidance, not case documentation.' The same report conceded per-output human validation is impossible; verification is batch-mode (expert curation before, monthly internal accuracy reports after, an annual corpus review against monthly platform builds), and an in-product warning — 'answers may not be fully accurate. Consult your supervisor as needed' — delegates per-query vigilance to staff. The whole loop sits inside one agency, in its first era without a court monitor in three decades; no independent audit or published accuracy data exists, and every outcome figure is a vendor claim from a customer story that never names CORA.

A nonprofit maps cities into small grid cells and models which places will see child maltreatment next, from crime, blight, and the built environment — never scoring a family, never touching a caseworker's screen. In Richmond, the builders' open-source model put ~70% of held-out maltreatment events in its top risk tier (double the density baseline) covering just 10% of the city — and their own published audit reported the model does not generalize across neighborhoods of differing racial composition, despite race being excluded from the features. New Hampshire is the one documented operational use: inside a 2018–2023 federal grant, coalitions targeted outreach and issued funded RFPs where the maps pointed. No independent outcome evaluation of any deployment exists, and the atlas's own structural reading names the quiet loop: attention goes where the map points, reporting can rise where attention goes, and next year's model trains on the records that attention wrote.