In this talk Stark argues that ML-driven science involving certain categories of data is fundamentally “conjectural.” Such “conjectural science” produces conclusions reliant on post-facto interpretation: partial inductive insights misinterpreted as widely applicable deductive truths. Elaborating on the notion of “conjectural science” as developed by Italian historian Carlo Ginzburg, he applies the history of the “conjectural sciences” to their contemporary instantiations in machine learning research and aiming to clarify some of the ways in which the use of ML systems is appropriate or inappropriate, not only practically but also conceptually. The appeal of AI-driven predictive statistical analyses, paired as they are with very large data sets to be analyzed, is that the conclusions drawn from that data about individual humans can be understood as measurable and repeatable through sheer quantity of data. However, he argues that AI techniques, when applied to data about humans and the social world, are conjectural science raised to its most acute form. Though AI analysis of human behaviors and societies aspire to the epistemological norms of “natural” science, in many cases, AI does not support such science as commonly understood.

Luke Stark is an Assistant Professor in the Faculty of Information and Media Studies at Western University in London, ON. His work interrogates the historical, social, and ethical impacts of computing and AI technologies.