Can AI have free will?

The debate over free will is old, and AI gives it a new test case. People ask the question in two moods: whether AI might already be choosing for itself, and whether it is “just following its programming”. Both turn on what free will requires.

Three views of free will

Libertarian free will requires that a choice not be determined by prior causes: at the moment of choosing, the agent could have done otherwise, with everything about the past held fixed. On this view a digital computer, whose next state follows from its current state and inputs, has no free will. Adding randomness does not help, since a random choice is not one the agent controls.

Compatibilist free will holds that freedom is compatible with causation. What matters is that a choice comes from the agent’s own reasons, values and deliberation, through a process the agent controls, without coercion. Philosophers in this tradition, such as Harry Frankfurt, Daniel Dennett, and John Martin Fischer and Mark Ravizza, give different accounts of the right kind of control, but none requires that the choice be uncaused.

Scepticism about free will holds that nobody has free will in the sense needed for deep moral responsibility, humans included. On this view the question about AI is easy, but only because the answer is no for everyone.

The argument that some AI has it

Christian List, in “Can AI Systems Have Free Will?”, argues for a compatibilist answer that some current systems may meet. On his view we should not look for mysterious properties, indeterministic algorithms or unpredictability, but ask whether we have good explanatory reasons to treat a system as an intentional agent: one with goals, a capacity to choose between alternatives, and control over the resulting action. If the best explanation of an AI agent’s behaviour attributes these to it, the agent has free will in a pragmatic and useful sense. The approach is inspired by Dennett’s intentional stance, and critics have questioned whether it separates functional autonomy from the kind of free will that grounds responsibility.

What is at stake

The question matters because free will is usually taken to be a condition of moral responsibility. If AI agents that book flights, write and run code, or trade on markets have free will in even a thin sense, it becomes tempting to say that they, rather than their makers, are responsible for what they do. Most philosophers resist that step: being free in a thin sense is not the same as being able to understand, answer for and be blamed for one’s actions. Where responsibility for AI harms actually goes is taken up in When an AI system causes harm, does responsibility disappear or move?

A related question is whether an AI can act for reasons at all without being conscious, which is discussed in Can a system act for reasons without being conscious? You can switch capacities on and off in the lab’s Agency Without Consciousness simulation and see which accounts of agency still count what you see as action.