Can AI Systems Have Free Will? Christian List's Pragmatic Test
Whether AI systems can have free will has a workable answer, and Christian List states it. His paper “Can AI systems have free will?” appeared in Synthese 206, article 115, in 2025 (DOI 10.1007/s11229-025-05209-x), and he posted the preprint to arXiv on September 14, 2026 (arXiv 2609.15407, CC BY 4.0). List is professor of philosophy and decision theory at LMU Munich and co-director of the Munich Center for Mathematical Philosophy, formerly professor of political science and philosophy at the London School of Economics and a British Academy electee. His framework says an AI system has free will if we have good explanatory reasons to view it as an intentional agent, with the capacity for choice between alternative possibilities and control over the resulting actions.
The question arrives late to a debate that has already covered moral agency and sentience for AI. List opens with that observation, and his framing is drawn from Daniel Dennett’s work, whose multiple drafts model this site examined in detail. The consequence is a test that runs on explanatory practice rather than on physics.
What the framework rejects
List clears the ground first. Three questions that dominate free will debates, in his account, do not belong in the AI case, and he argues they are the wrong questions for any case.
| Question List sets aside | The standard form | Why he rejects it for AI |
|---|---|---|
| Is there a mysterious property? | Look for an extra ingredient that makes will free | Explanation should not wait on an unknown essence |
| Are the algorithms indeterministic? | Free will requires an open physical future | Determinism does not block the relevant kind of choice |
| Is the system unpredictable? | Freedom means an outside observer cannot predict | Predictability is the wrong diagnostic for agency |
The three rejections share one structure. Each looks for a feature below or beside the level at which agency operates, a hidden substance, an open physics, or a blindfolded observer. List’s replacement stays at the level where explanation already works. We explain human action by treating people as agents with reasons, options and control. The question for an AI system is whether the same explanatory practice earns its keep.
What the replacement test asks
The replacement has three components, and each is an ordinary explanatory question rather than a metaphysical one. Is the system well modeled as an intentional agent, something whose behavior is explained by what it believes, wants and chooses? Does it hold a capacity for choice between alternative possibilities, in the sense that its deliberation connects to which option gets realized? Does it control the resulting actions, so that its states, not external quirks, settle what it does?
A positive answer on all three yields free will in what List calls a pragmatic and diagnostically useful sense. The word pragmatic is doing precise work. The verdict does not certify a metaphysical status. It licenses a form of explanation, and it carries diagnostic value because it tells engineers, users and regulators which systems their intentional stance actually fits.
For contemporary AI the test cuts across familiar ground. A language model queried through an API is a weak candidate on all three components, since its behavior is a function of prompts and weights with no persistent deliberation. A deployed agent with memory, goals and a planning loop is a stronger candidate on the intentional-agent component, and the open question is whether its choice and control components hold. List’s test gives that disagreement its terms.
Comparison to The Consciousness AI
The framework connects to the site’s coverage at two joints, and it diverges from a third. The first joint is Dennett. List’s intentional-agent test is compatibilism carried to machines, and Dennett drew the same consequence for artificial systems from his own theory, that a system with the right parallel organization has whatever free will and consciousness there is to have. The second joint is agency measurement. The active inference phenotyping work on agency in AI systems measures exactly the kind of self-directed organization List’s test responds to, and the autonomous agency scale applied to six contemporary systems scores the active and idle bands of the same boundary.
The divergence is deliberate. List’s paper addresses free will and explicitly leaves sentience aside, which puts it outside the consciousness verdicts this site tracks. On our position, consciousness is an emergent property of organized physical systems, and free will as List defines it is an explanatory attribution that can come apart from experience. A system could satisfy the intentional-agent test without phenomenal consciousness, and a phenomenally conscious system could fail it. The value of the paper for the consciousness debate is that it separates the two questions cleanly, which stops the field’s free will intuitions from smuggling consciousness verdicts in through the back door.
What the framework settles and what it leaves open
The framework settles how to talk about AI free will, and it leaves two things open. Whether any current system passes the test is an empirical question List does not attempt, and the paper’s own examples stop at the framework. The deeper open question is the relation between the pragmatic attribution and the metaphysical one. List argues the pragmatic sense is diagnostically useful, which is a claim about explanation. Whether it also tracks something real in the system is the residual metaphysical question, and his Dennettian answer is that the pragmatic use is as real as the concept needs to be.
For the site’s readers the useful move is the separation. The compatibilist emergence proposal from Blackburne and colleagues treats a form of emergence as necessary but not sufficient for consciousness, and List treats agency as sufficient for a pragmatic free will verdict without touching consciousness. Both papers tidy conditions that the AI consciousness debate tends to run together, and the flagship field survey shows the field still needs that tidying.