Abstract
Big data are used in child welfare in several ways. By linking previously disparate sources of data into large datasets, algorithmic risk models assign risk scores to new reports to assist with decision-making. Implemented in multiple jurisdictions to assist with triaging reports, promoters claim they increase accuracy, while critics argue risk scores based on previous data entrench social biases, dehumanize practice, and justify the hyper-targeting of resources within neoliberal states. Nor are these risk scores necessarily more accurate than humans, because child protection data underpinning the algorithm do not reflect child abuse incidence, false positives are high, and feedback loops in the data can distort predictions. Other ethical concerns are that decision-making is non-transparent, and that decisions are fundamentally unfair, because they are based on statistical similarity to a group, rather than individual actions. Mitigation of these concerns can only be partial, but suggestions are the use of design justice processes during tool development and use with family consent.