UC Berkeley

This dissertation is about the purpose of artificial intelligence (AI) research. New learning algorithms, scales of computation, and modes of sensory input make it possible to better predict or simulate decision-making than ever before. But this does not tell us whether or how AI systems should be built. In fact there is much anxiety about how to build AI applications in ways that respect or enact the decision criteria of existing human institutions. But instead of how to better predict or protect how we decide things, my research question is: how can AI tools be used to reorganize the choices we make about how we want to live together? Answering this question requires investigating the conditions under which deliberation is possible about the systems being built—their models, their real-world performance, and their effects on human domains. These three modes of deliberation are philosophically outlined in the introduction and named as sociotechnical specification, normative cybernetics, and machine politics.

The first chapter pursues sociotechnical specification in the context of routing algorithms for autonomous vehicle (AV) fleets. It asks what it would mean to relate this emerging transportation model to the other legacy systems adjacent to the travel domain. It sketches proxies in terms of “known unknown” features of the driving environment. These would need to be monitored and serve as targets for optimization in order for the performance of the AV fleet to be considered to be robustly good. The second chapter pursues a normative cybernetics of AVs in terms of a sustained internal critique of reinforcement learning (RL). This introduces new policy questions, whose answers would correspond to types of feedback between the behavior of AV firms and civil society or state organizations. The third chapter outlines the elements of machine politics in terms of concepts borrowed from contemporary analytic philosophy. Ruth Chang’s notion of parity is mobilized to demonstrate the possibility of domain deliberation at different stages of AI development. This comprises a critique of existing schools of thought, represented here in terms of epistemicism (the notion that the structure of human activities can be passively learned and observed) and ontic incomparabilism (the notion that human activities cannot be organically modeled or developed by means of AI). The three types of feedback that are produced through active developmental inquiry are presented in terms of featurization, optimization, and integration, all of which comprise the structural choices at stake in machine politics.