Oregon's Safety at Screening Tool — a Child-Welfare Risk Algorithm Retired in 2022 After Its Allegheny Model Showed Racial Disparity

From 2018 Oregon's Department of Human Services used the Safety at Screening Tool, which generated numerical scores estimating the likelihood a child would end up in foster care or be investigated again, to help hotline workers decide which child-abuse and neglect reports should be investigated. Workers kept discretion over the final decision. The tool was inspired by the Allegheny Family Screening Tool in Pennsylvania.
In April 2022 the Associated Press reported Carnegie Mellon University research on Allegheny's algorithm: it would have recommended that about two-thirds of Black children be investigated, compared with about half of all other children. Oregon said its tool included a 'fairness correction'. In May 2022 DHS told staff it would stop using the algorithm by the end of June 2022.
DHS's stated reason was that the predictive tool could not be incorporated into its new Structured Decision Making model, which relies on family-specific information; officials also referred to equity and continuous quality improvement. Senator Ron Wyden publicly criticised the use of 'untested algorithms' in decisions about families. No independent evaluation of Oregon's own tool has been published in the sources reviewed.
Lesson: agencies adopting imported predictive models should require local, published validation and disparity testing before deployment, and plan an exit path.

Read the full analysis: https://www.npr.org/2022/06/02/1102661376/oregon-drops-artificial-intelligence-child-abuse-cases