Abstract
Current debates on fairness in algorithmic decision-making tend to leave several questions unasked. Do the decisions rest solely on predictions, or also on non-predictive criteria? What are the differing effects of noise in the data as opposed to bias in the data? And is the social good being distributed by the decision genuinely desirable for the applicants receiving it? This talk argues that these questions expose a close relationship between fairness and the limitations of what a decision-maker can actually know, and considers what affirmative action means once that relationship is taken seriously.
On May 12, 2021, I gave an invited talk at the Carl Friedrich von Weizsäcker Colloquium of the University of Tübingen, held online.
The talk asked what the algorithmic fairness debate tends to leave out: whether decisions rest on predictions alone or also on non-predictive criteria, how noise in data differs from bias in data in its fairness consequences, and whether the good being allocated is one the applicants actually want. Taken together, these point at a link between fairness and the limits of knowledge available to a decision-maker.
A recording of the talk is available on YouTube.
Sandro Radovanović
Assistant Professor at University of Belgrade
My research interests include machine learning, development and design of decision support systems, decision theory, and fairness and justice concepts in algorithmic decision making.