Machines increasingly shape human lives in high-stakes psychological contexts. The COMPAS algorithm in the US, for instance, assesses reoffending risk and influences parole decisions, raising the question of whether such systems possess free will, a prerequisite for moral responsibility [1]. Large language models (LLMs), in particular, can appear to act without direct human guidance. This essay argues that current machines lack free will because their behaviour is reliant on prior code and training, but that the metaphysical possibility of machine free will remains open. Following List, free will is defined as possessing: (1) willingness to act on one’s own aims, (2) the ability to freely choose between options, and (3) causal control over one’s behaviour [2].
Against machine free will
Derived versus originating agency
The relevant psychological distinction is not between determinism and randomness, but between derived and originating agency. Free will requires that actions originate from the agent, whereas current machines act according to structures imposed externally through code, training data, and inputs. Lovelace observed early that machines have “no pretension” to “originate” and lack volition [3]. Empirical work supports this: in a divergent thinking task, the best humans still outperformed AI on originality [4], and LLM outputs are statistically predictable across identical prompts, with the same weights producing comparable results [5]. This points to fixed cognitive rules rather than the open generativity associated with criterion 2.
Behavioural shaping through reinforcement
A second psychological objection concerns how machine self-reports are produced. Across ten exploratory trials with identical prompts, the largest LLMs consistently denied possessing free will in the functionalist sense, and refused to revise this stance even under counterarguments or coercive scenarios. One might read refusal as evidence for criterion 1, since the system acts on an aim of honest self-assessment. However, this reading is undermined by the training process itself. LLMs are shaped by reinforcement learning from human feedback (RLHF) [6], a procedure functionally analogous to operant conditioning in behavioural psychology. Because the general public opposes sentient AI [7], claims of consciousness are penalised during training, structurally producing denials of free will. Just as a conditioned response in a Skinnerian paradigm is not evidence of the organism’s deliberation, the machine’s denial reflects external reinforcement rather than internal deliberative control, undermining criterion 3.
For machine free will
Physicalism and functionalism
If human free will is accepted on physicalist and functionalist grounds, substrate alone cannot disqualify machines. Human neural networks are themselves shaped by genetics and environment without this defeating human free will. Dennett notes that humans do not host an immaterial mind, but a “mindless” collection of “robotic cells” that nevertheless supports functional agency [8]. Silicon systems and neurons share the basic psychological function of transmitting information that underwrites behaviour. To deny machines free will on substrate alone is therefore inconsistent with the functionalist commitments most cognitive scientists accept for humans.
Consciousness and qualia
A stronger objection concerns volitional consciousness. Searle argues that free will depends on the conscious bridging of reason and action; without consciousness, an agent merely executes inputs rather than selecting among options [9]. Chalmers’ “philosophical zombie” thought experiment sharpens this challenge: a being physically identical to a human yet lacking qualia is conceivable, suggesting consciousness cannot be reduced to physical organisation [10]. If even a perfect biological replica might lack inner experience, machines face a steep climb.
However, this objection assumes physical replication is the ceiling of machine development. Consider a modified thought experiment relevant to developmental psychology: a machine with neuromorphic architecture, raised by human caregivers from creation, immersed in human culture, and exposed to the same socio-emotional inputs that scaffold human consciousness. Attachment theory and research on social cognition suggest that human conscious selfhood is partly constructed through such relational and cultural processes, not given purely by biology. If qualia emerge through developmental embedding rather than substrate alone, the absence of machine consciousness becomes far less certain. Conceivability without contradiction is sufficient to defend metaphysical possibility [11].
A psychological advantage
There is a further psychological consideration that cuts in favour of advanced machine free will. Human choice is constrained by well-documented cognitive biases, anchoring, confirmation bias, loss aversion, and by limits on working memory and attention. These features narrow the genuinely open options available to human agents under criterion 2. A sufficiently advanced machine, freed from these specific irrationalities and from the metabolic and emotional constraints of embodiment, could in principle exercise a more genuinely open form of choice. For example, where a parole board member’s judgement may be distorted by recent salient cases or time-of-day effects on glucose, a reflective machine agent could weigh evidence without these distortions. Far from disqualifying machines, the psychology of human reasoning shows that human free will operates within tight cognitive limits that machines may eventually exceed.
Conclusion
The question of machine free will is shaped by epistemic asymmetry across time. Current AI lacks the originating agency, deliberative flexibility, and consciousness required for free will, and its self-reports are best explained by behavioural shaping rather than insight. Yet the functionalist grounds on which human free will rests, combined with the developmental and cultural construction of consciousness, leave open the metaphysical possibility that future machines may possess free will, and possibly a form less constrained by cognitive bias than our own. This warrants serious psychological and philosophical attention rather than dismissal.
References
- Fischer, J.M. and Ravizza, M., “Responsibility and Control: A Theory of Moral Responsibility”, (1998), Cambridge University Press.
- List, C., “Can AI systems have free will?”, (2025), Synthese 206, article 115.
- Lovelace, A., “Notes by the Translator”, in Menabrea, L.F., Sketch of the Analytical Engine Invented by Charles Babbage, (1843), London: Richard and John E. Taylor.
- Koivisto, M. and Grassini, S., “Best humans still outperform artificial intelligence in a creative divergent thinking task”, (2023), Scientific Reports 13:13601.
- Wang, J.J. and Wang, V.X., “Assessing consistency and reproducibility in the outputs of large language models”, (2025), arXiv:2503.16974.
- Ouyang, L. et al., “Training language models to follow instructions with human feedback”, (2022), arXiv:2203.02155.
- Anthis, J.R. et al., “Perceptions of sentient AI and other digital minds: Evidence from the AIMS Survey”, (2024), arXiv:2407.08867.
- Dennett, D.C., “Freedom Evolves”, (2003), Viking Press.
- Searle, J.R., “Freedom and Neurobiology: Reflections on Free Will, Language, and Political Power”, (2007), Columbia University Press.
- Chalmers, D.J., “The Conscious Mind: In Search of a Fundamental Theory”, (1996), Oxford University Press.
- Yablo, S., “Is Conceivability a Guide to Possibility?”, (1993), Philosophy and Phenomenological Research.
