Agency and Imitation

Abstract AI provides a rich source of challenging examples for theories of agency because it allows relevant features to be relatively freely combined. This paper considers agency in sequence models: models that have been trained to predict the next item in a sequence. Training models on sequences of observations and actions can yield systems that use reinforcement learning in context, as well as ones that merely imitate agents. Investigating AI systems from this group suggests a requirement that agents must learn about the environment in part from their own interaction with it. It also suggests a distinction between two alternative approaches to understanding agency.