Domain Atlas / Child welfare & family services
Sistema Alerta Niñez (Chile)
Work with this case in the PAN Lab ↗
In a 2019 proof of concept, Chile's Sistema Alerta Niñez risk models reached test-set AUC of roughly 0.88 to 0.95 for a two-year outcome — a child's separation from family or contact with child-protection programs — using 280 administrative variables per child; the deployed operational model's real-world performance was never publicly disclosed.[3]
What happened
Chile's Sistema Alerta Niñez scores each child's risk of a future rights violation and produces a prioritized list that Local Childhood Offices (Oficinas Locales de Niñez) use to decide whom to reach out to first. It grew out of President Piñera's 2018 "Gran Acuerdo Nacional por la Infancia"; the predictive model was built in 2018–2019 by a consortium of GobLab at Universidad Adolfo Ibáñez and the Centre for Social Data Analytics at Auckland University of Technology, and was implemented and maintained by the Chilean firm Actis. The developers explored several methods and selected LASSO regression; in a 2019 proof of concept the models reached test-set AUC of roughly 0.88 to 0.95 for a two-year outcome — a child's separation from family or contact with child-protection programs — drawing on 280 administrative variables per child from benefit, education, health, child-protection, crime and census records. Officially the score is "one more input," always subordinate to the judgment of OLN professionals. The tool was piloted in 12 communes (about 3,354 children served by August 2020), and the OLN network later expanded toward a majority of the country's communes. The deployed model's real-world performance was never publicly released, and an external algorithmic-bias audit — reported to be funded by the Inter-American Development Bank and conducted by the consultancy Eticas — was carried out but its criteria and results were never made public.
The sociotechnical reading
Most cases in this Atlas turn on the output side — who acts on a score and who can correct it. Sistema Alerta Niñez turns on the input side. The data that builds the ranking was gathered for another purpose entirely: families supplied it to qualify for social benefits through the household registry, and the predictive-modeling use was recast as ordinary "targeting" (focalización), so the people scored were never told a risk ranking existed, could not opt out, and were not consulted. In map terms the governance failure lives on the pathway from records into the model, not on the pathway from model into decision — a consent and purpose-limitation gap upstream of any override. Two controls that might have bound the system were built but never turned on: the independent bias audit was commissioned and then withheld, the operational model's performance was never published, and the field knowledge OLN staff gather was never fed back to the model. An audit whose results no one may see is not a control, and a score whose accuracy is never disclosed cannot be contested. The lesson is that transparency at the input and evaluation boundaries — not accuracy at the output — is where a tool of this shape is governed or left ungoverned.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library. The model organization for this case can be stress-tested in the PAN Lab.