[Postponed] Dr. Louise Amoore Lecture on Machine Learning Politics

The idea that a ‘good’ machine learning model is one that can generalise to new situations has a long history. Even Turing’s 1950s accounts of machine intelligence referred to what he called a “spring of action” that exceeded the programming of explicit rules. By 2012, when the Turing Laureate Yoshua Bengio sets out the guiding principles for unsupervised machine learning, the ‘good’ model is rendered normatively as having the capacity to exploit the unknown structure in data. Here, that which is unknown and unencountered is re-cast as a positive force to be harnessed in machine learning. It is a machine learning logic that has simultaneously become pervasive in the contemporary governing of societies – how the unknown structure of health data, policing data, pandemic data, immigration data, might yield the patterns and features that make interventions possible. The combinatorial possibilities of deep learning models reimagine the contingencies of the world as a field of political possibility. When Bengio proposes that deep learning algorithms “discover good representations” in data distributions, I propose that this logic powerfully generates a politics of discovering good representations of a social distribution. Thus, to deploy large language models (LLMs) or transformer models in the social world is never only to instrumentally bring a tool into use, but rather it brings into being a specific political means of picturing and knowing the world.