When the algorithm decides for vulnerable communities

When the Algorithm Decides: AI Safety, Gender-Based Violence, Coercive Control and CALD Communities AI safety frameworks are built on assumptions about who uses AI and who is harmed by it. For women and girls experiencing family violence, particularly those from culturally and linguistically diverse communities, those assumptions are often wrong, and the consequences are life-threatening. This presentation draws on active work within the IEEE Standards Project on User-Centred Principles for AI Used in Evaluating Family Violence to examine what AI safety must mean when the deployment context is coercive control, lethality risk assessment, and algorithmic custody decisions. For CALD communities, the stakes compound: language barriers, cultural stigma, immigration status, and systemic distrust mean that poorly governed AI does not merely fail, it actively re-harms. We argue that survivor-centred design principles are not an addition to AI safety, they are its foundation. Safe AI, in this context, means explainable, bias-audited, human-supervised systems designed with the most vulnerable users at the centre. With Australia deploying these tools now, the standards must catch up.