A presentation by Dr Brent Mittelstadt, Director of Research at Oxford Internet Institute. This event is part of the Fundamental Rights Research Colloquium hosted by the Centre for Fundamental Rights.
Western societies are marked by diverse and extensive biases and inequality that are unavoidably embedded in the data used to train machine learning. Algorithms trained on biased data will, without intervention, produce biased outcomes and increase the inequality experienced by historically disadvantaged groups. Recognizing this problem, much work has emerged in recent years to test for bias in machine learning and AI systems using various fairness and bias metrics. In this talk, Brent will introduce two ways in which this work in fairML can and does result in morally and legally problematic decision-making. First, he will introduce the concept of “bias preservation” as a means to assess the compatibility of fairness metrics used in machine learning against the notions of formal and substantive equality as embedded in non-discrimination law. Second, he will introduce the essential elements of a new piece of work on the phenomenon of levelling down in fairML, and discuss future directions for this work.
Brent Mittelstadt is the Oxford Internet Institute’s Director of Research, an Associate Professor and a Senior Research Fellow. He also coordinates the Governance of Emerging Technologies (GET) research programme which works across ethics, law, and emerging information technologies. He is a leading data ethicist and philosopher specializing in AI ethics, professional ethics, and technology law and policy. In his current role, he leads the Trustworthiness Auditing for AI project, a three-year multi-disciplinary project with the University of Reading cutting across ethics, law, computer science, and psychology to determine how to use AI accountability tools most effectively to create and maintain trustworthy AI systems.
Prior registration is required. Registered attendees will receive the dial-in details as well as a draft paper, on which the presentation is based, via e-mail prior to the event.
