Project Summary. In acute psychiatry, patient aggression and violence towards self or others are of paramount concern. These behaviors can negatively impact the well-being of patients and staff. Further, common practices used to manage patient aggression, including chemical and physical restraints and seclusion, are coercive practices that can be (re)traumatising. Many acute psychiatric settings employ structured rating scales to assess early warning signs of aggression, such as agitation. While structured scales can be effective early warning and prevention tools, high false positive rates are concerning (eg, 23-50% of patients rated as high risk do not go on to become violent or aggressive). Given the limitations of these scales there are efforts to enhance these assessments with Artificial Intelligence (AI), by training machine learning models on patient data to generate predictions about who is likely to become violent or aggressive. However, our work demonstrates that these models are more likely to overestimate this risk for racially marginalized, Indigenous and and structurally marginalized patients. Thus, such applications of AI are unlikely to enhance clinical practices and perpetuate existing inequities. To prevent these discriminatory harms we propose an alternative approach. Rather than use AI to identify individuals at risk of violence, the overall goal of our project is to use AI to examine the features that contribute to false positive predictions and then co-design a clinical decision support system designed to provide additional information designed to support equitable and patient-centered care. To address this aim, our project combines qualitative and quantitative approaches to assess the potential of using AI to advance equitable and compassionate risk assessment in acute psychiatry.