WP 1
AI and the fundamentals of agency
What does it mean to act? And in what sense does an AI system act? This work package builds a philosophical framework for agency drawing on the philosophy of biology: the notion of scaffolding, whereby complex functions emerge from enabling structures, serves as a model for embedded agency — that of AI systems, whose capacities depend on human-designed infrastructures and training regimes.
Lead: Philippe Huneman (IHPST) Contributors: Maud van Lier, Matteo Mossio, Francesca Poggiolesi, Alberto Naibo, Pierre Wagner, Catherine Pelachaud, Mehdi Khamassi, Raja Chatila, Alexandre Gefen, post-doctoral researcher (IHPST)
Questions
In philosophy, agency is mostly related to humans as deliberative rational agents. Yet some biological phenomena seem to require attributing agency to organisms deprived of consciousness: agency may thus be correctly attributed to non-human systems. The work package explores whether this framework of environmentally scaffolded agency can be used to understand AI systems, and whether biological agency offers a useful model for thinking about forms of AI agency that are non-human yet genuinely agential.
Three subtasks
- The fundamentals of agency — goal-directedness, teleonomy, self-organisation; the conditions of a “unity of purpose” (units of selection, holistic organisation, closure of constraints) and their role in scaffolding.
- Formalising scaffolded forms of agency — how forms of agency are built through layers of organisation and extended by interaction with pre-existing structures, agents or environments.
- The concept of embedded AI agency — measuring AI systems on a scale of agency, conceiving human–AI dyads as embedded agents, and deriving norms for human–AI systems from explainable AI (XAI).