Abstract: Prominent debates in AI ethics contend that biases are one of the main problems to be eradicated to design fair AI systems, leading the community to develop many technical and non-technical methods focused on eliminating them to create fairer AI tools. However, such methods are shown very often to be inadequate or counterproductive for the design of fairer AI systems, failing in fostering social justice as fairness through AI, especially in the domain of healthcare. As I will argue in this talk, one of the main problems is a procedural and merely mathematical understanding of fairness, which leads a passive understanding of biases just as mere errors and technical bugs in AI systems, as well as of the related solutions mainly aimed at their simple removal. To address this problem, this talk will go beyond mathematical and procedural conception of fairness, drawing conceptual insights from moral philosophy, to show what substantial fairness means and how to achieve it by leveraging responsible multi-agent architectures and explanations. Particularly, through a case study on gender and ethnic biases in healthcare AI, this talk will show how a substantial account fairness helps harnessing bias in developing multi-agent architecture-based AI systems to decode patterns of unfairness and, drawing on ethical theories of affirmative action, generate novel compensatory design actions for the development of truly fair healthcare AI systems, capable of addressing longstanding unfair inequalities in healthcare, therefore promoting through AI better and more just healthcare ecosystems.

Abstract: This talk will summarize some of our recent conceptual work on responsible autonomy, highlighting the opportunities and challenges for explanations. This work steps back from computational models of morality to apply insights from philosophy and social psychology to responsible autonomy. One challenge is to understand when responsible autonomy involves respecting of deviating from norms. We show how Habermas's notion of objective, subjective, practical validity criteria can be used to understand norm deviation. We apply case law from the US, UK, and Canada as a source of empirical knowledge about when norm deviations are legitimate. We apply the Habermasian framework as a basis for conceiving of consent. We adapt a model of trust based on the components of ability, benevolence, and integrity to show how trustworthiness involves, again applying case law as a source of empirical intuitions about how to evaluate AI agents. These models can provide a basis for effective and meaningful explanations as integral to responsible autonomy.