Winning, Jason (2019). The Mechanistic and Normative Structure of Agency. Dissertation, University of California San Diego.

Abstract:

I develop an interdisciplinary framework for understanding the nature of agents and agency that is compatible with recent developments in the metaphysics of science and that also does justice to the mechanistic and normative characteristics of agents and agency as they are understood in moral philosophy, social psychology, neuroscience, robotics, and economics. The framework I develop is internal perspectivalist. That is to say, it counts agents as real in a perspective-dependent way, but not in a way that depends on an external perspective. Whether or not something counts as an agent depends on whether it is able to have a certain kind of perspective. My approach differs from many others by treating possession of a perspective as more basic than the possession of agency, representational content/vehicles, cognition, intentions, goals, concepts, or mental or psychological states; these latter capabilities require the former, not the other way around. I explain what it means for a system to be able to have a perspective at all, beginning with simple cases in biology, and show how self-contained normative perspectives about proper function and control can emerge from mechanisms with relatively simple dynamics. I then describe how increasingly complex control architectures can become organized that allow for more complex perspectives that approach agency. Next, I provide my own account of the kind of perspective that is necessary for agency itself, the goal being to provide a reference against which other accounts can be compared. Finally, I introduce a crucial distinction that is necessary for understanding human agency: that between inclinational and committal agency, and venture a hypothesis about how the normative perspective underlying committal agency might be mechanistically realized.

I have not had a chance to read this, and my time is rather constrained at the moment so it's unlikely I will, but I stumbled across this, and it piqued my interest. Better understanding of agency appears important to the success of many research programs in AI safety, and this abstract matches enough of the pattern of what LW/AF has figured out matters about agency that it seems well worth sharing.

Full text of the dissertation here.