Agency

Agency is the capacity to act: to do something, rather than merely have something happen, for reasons that are in some sense one’s own. The standard account, owed largely to Anscombe and Davidson, picks out action by its relation to reasons. An event is an action when it is intentional under some description, and on Davidson’s version when it is caused in the right way by the agent’s beliefs and desires.

Agency comes in grades. A thermostat responds; a dog acts; a person acts and can ask whether she should have. Artificial systems now occupy an unfamiliar rung, and a more useful question than whether a system “has agency” is which of the capacities bundled under the word it has: goal-directedness, sensitivity to reasons, the ability to revise a plan, the standing to be held to account. Floridi and Sanders argued that artificial agents can be sources of morally significant action, and accountable for it, without mental states, while reserving responsibility for agents of another kind.

My own arguments are built so as not to depend on how this is settled. Whether a language model transmits knowledge turns, on my account, on what governs its production of assertions, not on whether it is an agent, a mind or a human being; the Tracking Minimum is framed so that a machine could satisfy it. The charge that the received view of testimony has been anthropocentric is often right, and an argument resting on agency or mentality would concede the diagnosis while denying the conclusion.

Agency returns in my political work from the other side. Consent presupposes an agent who could authorise or refuse. Where an institution produces a person’s operative standing as an entitlement-bearer, that standpoint is not available, and the question becomes what the institution owes her rather than what she has agreed to (see constitutive dependence). Who counts as an agent decides who can be held responsible, and the question is now a practical one.