Jonas Wahl (German Research Centre for Artificial Intelligence (DFKI) and Prof. II at UiB)

The trouble with individualized counterfactuals

Abstract

Modeling counterfactual outcomes has become increasingly popular in contemporary machine learning with applications from medical decision-making to machine ethics. At the same time, the use of counterfactuals in statistics has been controversial among statisticians for many decades. In this talk, I will review the debate on counterfactuals in statistics, including its most recent iteration revolving around counterfactual notions of harm, in light of current developments in artificial intelligence. In particular, I will argue that the range of circumstances in which counterfactual estimation can be convincingly justified is narrow and that there is tension between common machine learning practices and counterfactual modeling assumptions.