Abstract: 
Artificial intelligence (AI) offers the potential to revolutionize practices across industries, professions and society. However, while AI has the potential to provide significant benefits, it also comes with potential for significant harm. To ensure that the benefits of AI are realised and any risks are mitigated, it is crucial to move beyond speculation and hype and develop an evidence-based understanding of the impacts of AI. There is, however, currently no consensus on how best to evaluate the impacts of AI. In this talk, I introduce Evidential Pluralism as a promising methodology for evaluating AI. By combining evidence of association and evidence of mechanisms, Evidential Pluralism, I argue, can help to capture the complexity, context sensitivity and rapidly evolving nature of the impacts of AI. This, in turn, can enable the development of justified and effective legal interventions.