<p>Artificial intelligence systems increasingly automate decisions traditionally made by human beings. Contemporary AI ethics, however, lacks a rigorous, quantitative method for evaluating the impacts of automation on human wellbeing. Existing approaches like impact assessment and AI alignment remain either too qualitatively descriptive or normatively shallow. This offers no principled basis for deciding among the ethical tradeoffs that AI deployments often demand. This paper presents a general method for quantitatively measuring the human impacts of AI: the Human Impact Scorecard. Our methodology makes two epistemic commitments. The first is humility: we do not purport to assert a theory of the good; we seek a pragmatic framework that can accommodate a wide range of normative commitments. The second is ecumenism: our method must be tolerable to a diverse range of perspectives, and portable across domains necessary for human flourishing. To satisfy these commitments, we adopt and extend Nussbaum and Sen’s capabilities approach as our normative foundation. The capabilities approach identifies a plural set of dimensions that together constitute a good life; they are each amenable to quantitative measurement. This in hand, we develop a two-step method: first, we identify the capabilities most relevant to a given domain of AI deployment; second, we operationalize those capabilities using established metrics that permit robust quantification; this produces a normatively grounded multicriteria matrix that can be used to broker a set of Pareto-frontier options. We illustrate this using a case study involving competing biometric authentication products, and by discussing the method’s implication for AI alignment and reward function design. We conclude by discussing the conceptual and normative challenges that remain.</p>