Making Machines That Make Us: How Machine Learning Research Shaped Human Capacity and Social Possibility

• In Event: Precision and Uncertainty in a World of Data - I

Abstract

Presently there exists a plurality of ideas within the computer science community, STS and legal scholarship, and the general public about (1) what is connoted by the terms “machine learning” and “artificial intelligence,” and (2) what the relationship is between these two fields of inquiry. Rather than resorting to surveys, ethnographic studies, or posthoc textbook definitions, this talk narrates the historical uses of “machine learning” in the research literature written in the 1950s to the present, paying particular attention to the early “machine learning” systems of the 1950s and 1960s as well as the heretofore unacknowledged diversity of “expert systems” from the 1970s to 1990s. Ideas of both human and machine “learning” within these communities was formed through a triangulation between specific research problems, speculative intellectual and professional ambitions, and specific material computing devices. Machine learning researchers operationalized human performance not by starting with observed human behavior but rather with the capabilities of particular computing devices to establish benchmarks that were subsequently used to define and evaluate *both* human and machine performance. In addition to providing a radically alternative version of AI history, this paper demonstrates how the strategies used in contemporary research on fairness and transparency (FAT) in machine learning largely follows strategies and epistemological modes of knowledge creation from very specific and largely unknown machine learning problems in the early 1950s. Closely attending to the epistemological innovations of these 1950s researchers allows us to identify the shortcomings of contemporary FAT literature and to postulate alternative approaches not subject to the same limitations.

Author

• Aaron Louis Plasek, Columbia University