Can AI systems have free will?

In plain words

People ask whether AI systems could be moral agents or conscious. This essay asks a different question: could an AI system have free will? The author says we should not start by hunting for some special inner ingredient, by checking whether the underlying algorithm is random, or by checking whether the system is hard to predict. Those tests do not separate people from machines.

Instead, the test is practical. Look at the system and ask whether the best explanation of what it does treats it as an agent. Does it act toward goals, based on something like beliefs and wants? Does it pick one option among several real alternatives? Does its own goal-driven state actually cause what it does? If the answer is yes, the system has free will in a useful, everyday sense.

The idea draws on Daniel Dennett's "intentional stance", which is the habit of explaining something by what it wants and believes. The essay's main point is that this habit, when it is the best explanation, is evidence that the thing is really an agent.

TL;DR

• Free will is defined as three jointly necessary and sufficient conditions: intentional agency, alternative possibilities and causal control §3, p. 5.
• The test for an AI system is whether it is best explained as meeting all three, not whether it is mysterious, indeterministic or unpredictable §4, p. 7; §6, p. 18.
• The paper gives no experiment or measurement; its verdict is a conditional argument, and no current system is assessed.

Problem

The paper addresses a gap in the AI debate: whether AI systems could have free will has received little attention, while moral agency and sentience have been discussed much more §1, p. 1. Free will matters because it is often treated as a precondition for responsibility §1, p. 1. The few prior works that discuss it mostly conclude negatively, because they use conditions the author considers too demanding §1, p. 2.

Two common worries are set aside. First, free will is not taken to need a metaphysical extra ingredient, such as a homunculus or a way to escape the laws of nature §4, p. 6. Second, the paper does not require the system to be unpredictable or to run on indeterministic algorithms §5.2–5.3, pp. 13–15.

Method

The method is an argument about criteria, not a computational procedure.

Definition of free will §3, p. 5. The paper uses a checklist of three conditions, following List (2019):

1. Intentional agency. The bearer is an intentional agent, meaning it acts in a goal-directed way on intentional states such as beliefs and desires.
2. Alternative possibilities. The bearer has different courses of action it could take.
3. Causal control. The bearer has relevant control over its actions, in the sense that its intentional states are the difference-making causes of those actions.

The paper assumes these conditions, "perhaps with some fine-tuning", are jointly necessary and sufficient §3, p. 5. Free will in this sense is a practical, third-person notion. It is not a claim about metaphysics §5.6, p. 17.

Application to AI §4, pp. 7–8. The paper proposes that a system has free will if and only if it meets the three conditions. The test is whether there are good explanatory reasons to view the system as meeting them §4, p. 7. An explanation that treats the system as an intentional agent is evidence that it is one. It is not a definition of agency §4, p. 7, fn. 6.

The paper's test for each condition, restated as a checklist:

# Test for free will in an AI system  [§4] [§4.2] [§4.3] [§6]
# Input:  system, described at the level of agency
# Output: True if the system counts as having free will

def has_free_will(system):
    # 1. Intentional agency: is the system best explained as a goal-directed
    #    agent with beliefs and desires?  [§4.1]
    if not best_explained_as_agent(system):
        return False

    # 2. Alternative possibilities: does the best explanation assume that the
    #    system faces a choice, evaluates options and picks one?  [§4.2]
    if not explanation_assumes_choice(system):
        return False

    # 3. Causal control: are its intentional states the difference-making
    #    causes of its actions, rather than only its low-level states?  [§4.3]
    if not intentional_states_make_difference(system):
        return False

    return True

The checklist is a restatement of the paper's argument, not an algorithm the paper gives. The paper says it has no operational test: the judgment is made by comparing explanations.

Condition 1, intentional agency §4.1, pp. 8–10. Belief and desire are read as representations of how things are and as a target state or ranking of states to achieve §4.1, p. 8, fn. 7. The paper cites work arguing that large language models may have beliefs and world models, and says it is largely an empirical question §4.1, pp. 8–9. High-level description is legitimate even when a low-level algorithmic description exists §4.1, p. 9. Explainable AI is treated as a reason to build systems whose behaviour is explained in agential terms §4.1, p. 10.

Condition 2, alternative possibilities §4.2, pp. 10–11. The paper gives a three-step pattern for intentional explanations: (1) the entity has a choice between options; (2) it considers or evaluates them from a goal-directed view; (3) it chooses one on that basis §4.2, p. 10. The paper argues that any decision-theoretic explanation of an AI system assumes such choices. Explanations of "why this rather than that" are contrastive, which fits systems that choose among alternatives §4.2, p. 11.

Condition 3, causal control §4.3, pp. 11–13. A mental state is a difference-making cause if (1) without it the action would not occur and (2) with it, in similar circumstances, the action would occur §4.3, p. 11. Control variables are the parameters that change outcomes of interest §4.3, p. 12. The paper argues that for AI, the question is whether high-level representational or goal states are better control variables than low-level algorithmic states. Where they are, the system has causal control over its actions §4.3, p. 12.

Degrees §4.3, p. 13. Systems can differ in the richness of their agency, the scope of their choices and the extent of their control. Judgments of free will may therefore come in degrees §4.3, p. 13.

Hyperparameters

Not applicable. The paper has no algorithm, model, training run or numerical parameter, and it contains no equations, tables or algorithm blocks §6 p. 18.

Results

Not applicable. The paper reports no experiments or benchmarks §6 p. 18. Its output is a set of conditions and a conditional conclusion: if an AI system is best explained as an intentional agent with alternative possibilities and causal control, then it has free will in the pragmatic sense §4.3 §6. The paper does not assess any named system §6 p. 18.

Paper vs code

No official code.

The arXiv abstract and PDF contain no code link, and the paper has no project page. A GitHub repository search for "free will AI list" returned 17 results, all unrelated to this paper or its author. The paper describes no software.

Pitfalls

• The test has no threshold. The paper does not say how strong an explanation must be before it counts as "good explanatory reasons". Its only comparison is to explanations that are "explanatorily clearly superior" to non-agential ones §5.4, p. 16. This is a judgment, not a measurement.
• Simple systems also pass in principle. The author accepts that a chess program or a thermostat could be described as a simple choice-making agent §5.4, pp. 15–16. The author resolves this by requiring that the agential description be clearly superior, so it is not attributed to them by default §5.4, p. 16. An implementer should not treat "passes the checklist" as "has free will" without this comparison.
• Explanation is evidence, not constitution. The paper uses an analogy: a driving licence indicates driving ability but does not constitute it §4, p. 7, fn. 6. Failing to find an intentional explanation is therefore not a proof that a system lacks agency.
• Determinism is handled at the agential level. The paper holds that the micro-level can be deterministic while the macro-level has alternative possibilities, because the relevant notion of possibility is conditional on the macro-state §5.2, p. 14. Its "agential possibility" follows List (2014, 2023) and Maier. This departs from Dennett (1984), who drops the requirement of alternative possibilities §5.2, p. 14; §3, p. 5.
• Which entity has free will. It is the system as a whole, the choice-making agent implemented by the hardware and software, not the computer as a physical device §5.1, p. 13.
• Predictability is not required. Predictable choices can still be free if they are made for intelligible reasons §5.3, p. 15.
• Reactive systems. Many AI systems act only when prompted. The paper says free will applies while the system is in a phase of choice-making agency, and it may be dormant between such phases §5.5, p. 16.
• Consciousness is not required. The paper treats the link to consciousness as a contingent matter. A system could have free will without any subjective experience §5.6, pp. 16–17. The paper keeps free will and the experience of free will separate §5.6, p. 17.
• Free will is not moral responsibility. Free will is necessary but not sufficient for responsibility, and the paper leaves the normative question open §5.7, pp. 17–18.
• Date discrepancy. The PDF header says it is a preprint of a paper published in Synthese (2025), and the title page says "This version: August 2025". The arXiv metadata lists 2026-09-14, and this card uses that date in published.
• Notation. The paper refers to "physical" and "functional" free will in the critique of Martela (2025). The author prefers to speak simply of free will §1, p. 2, fn. 3.

Lineage

• Builds on: List (2019) for the three-condition checklist and its application to AI §3, p. 4; §4.2, p. 10. Dennett (1984, 1987, 2003) for the intentional stance and "free will worth wanting" §3, p. 4; §4, p. 7.
• Related: Maier (2023) argues that AI systems can be choice-making agents to which decision-theoretic models apply §1, p. 2. Martela (2025) defends functional free will in generative agents. The paper differs from Martela by not separating physical and functional free will p. 2, fn. 3.
• Compared against: Floridi and Sanders (2004), Krausová and Hazan (2013), Sanchis (2018) and Blum and Blum (2022) are discussed as more demanding or as relying on unpredictability §1, p. 2; §5.3, p. 15.
• Analogy: the paper argues that corporations and other organised collectives can satisfy the same three conditions §6, pp. 18–19.
• Superseded by: not stated.