Having sent off the final version of my concepts book at the very end of 2023, in 2024 I finally had the space to start fulfilling a long-held aspiration to do some more reading around the topic of agency. Of course, teaching and working in philosophy of mind already makes the topic familiar, but I had long been conscious that the literature was deep and my understand was not.
My earlier book on representation was confined to subpersonal cases. One of the things I noted about personal-level representations is that they occur in a sphere were the person operates as an agent. This seems like an important difference but I am still not clear about its implications for how we should think about representational content. In my subsequent concepts book I aimed to put the thinker, as agent, at the centre of the picture. People think with their concepts – it is something they do. Thinking performed by the agent has various psychological attributes. It calls on what psychologists call executive functions: inhibition, working memory, goal-driven direction of attention, and so on. Metacognition is involved. This goes some way towards characterising thinking as an active process but falls a long way short of saying what it is to be an agent.
Every interesting question is connected to lots of others. Any research project is forced to treat some issues as out of scope. Still, I was worried about the risk of pushing a bump in the carpet conveniently out of sight. So it has been a pleasure to start reading into the topic some more. There are actually at least four relevant bodies of literature. There is of course a rich tradition in philosophy: philosophy of action, moral psychology and the metaphysics of free will. There is also newer work in philosophy of agency that incorporates empirical findings, for example about sense of agency, implementation intentions, habits and skills, and motor control. Second, there is a mass of work in the cognitive sciences. Frustratingly for us, almost nowhere is it collected together under the term ‘agency’. Highly-relevant findings are found in fields like decision-making, reinforcement learning, emotions, and motor control. Third, there is another whole rich tradition of work on biological ‘agency’: the kind of goal-directedness exhibited by even the most simple single-celled biological organisms. Fourth, there is an emerging body of research on agency in relation to artificial intelligence.
My interest in AI and computational models of the mind goes right back to my PhD. Having worked then on connectionist systems, it was heartening when, in 2012, they started to prove really useful. I followed, with interest, developments in deep neural networks, word-embedding models, and then the transformative Transformer architecture. In 2021 this was given a boost for me by the Deep Learning online reading group organised by Raphaël Millière, and more recently by connections to researchers at Deepmind in London. The power of LLM-based systems has made it seem that we may be on the verge of being able to build AI agents. But what is it for an AI system to be a genuine agent? It has been heartening to find that machine learning researchers are keen to talk to philosophers about this. Conversely, thinking about how to build the capacity for agency in artificial systems is throwing up new questions and helping to sharpen our understanding of what agency really involves.
My reading has been driven by curiosity, plus lots of discussions with helpful colleagues. But curiosity often turns into ideas, and ideas eventually become papers. My first paper in this new line of research has just been accepted at Philosophy and the Mind Sciences. It shows, I hope, that thinking about agency in the AI context can give us useful insights that apply to agency more generally.
