{"id":"14c20a6a-0054-4239-86b1-d8a52bb91cc7","arxiv_id":"2609.15407","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"An AI system can be said to have free will if it is best explained as an intentional agent that chooses between alternative possibilities and causally controls its actions.","lead":"This paper argues that whether an AI system has free will should be assessed by checking three conditions: intentional agency, alternative possibilities, and causal control. If an AI system meets these conditions, it qualifies as having free will in a pragmatic, useful sense, just like humans do.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The 'explanatorily clearly superior' criterion for applying the three conditions to AI systems is not operationally defined, leaving the central claim's diagnostic utility ambiguous.","rationale":"The reader's weakest assumption was the sufficiency of the three conditions for free will, which is a definitional issue. My concern is different: even if one accepts that definition, the application of the conditions to concrete AI systems relies on an underspecified 'best explanation' criterion. The paper anticipates the over-ascription problem but provides no systematic way to decide when the intentional stance is genuinely indispensable versus merely useful. This is a load-bearing gap because the central claim's practical value hinges on making such judgments transparently. I recommend a CONDITIONAL verdict: accept the philosophical framework in principle, but require a more explicit operationalization of the 'clearly superior explanation' test before the criteria can be considered diagnostically useful. The reader's ACCEPT with high confidence may be too optimistic given this unresolved methodological issue.","tokens_in":18846,"tokens_out":2380,"duration_ms":22988,"concrete_test":"Provide a detailed case study of a specific AI system (e.g., GPT-4 operating autonomously on a long-horizon task) and explicitly apply the proposed criteria: (1) articulate both an intentional-stance explanation and the best available non-intentional mechanistic explanation; (2) state precisely which criteria (e.g., predictive power, parsimony, fertility, ability to handle counterfactuals) are used to judge which explanation is 'clearly superior'; (3) show that the judgment does not reduce to a purely pragmatic preference. If the criteria remain ambiguous or the intended superiority judgment cannot be made in a principled, reproducible manner, the paper's central diagnostic claim is significantly weakened.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (an AI system has free will iff we have good explanatory reasons to view it as an intentional agent with alternative possibilities and causal control) relies crucially on the notion of 'explanatorily clearly superior' explanations (Section 5.4). The paper acknowledges the risk of over-ascription (e.g., thermostats) and attempts to block it by requiring that the intentional stance be 'explanatorily clearly superior' to non-intentional alternatives. However, no principled criteria are given for what makes one explanation clearly superior. The paper merely appeals to practical feasibility and simpler descriptions, but these are context-dependent and potentially subjective. Without an operationalized test for when the intentional stance is indispensable (rather than merely convenient), the framework cannot reliably distinguish genuine candidates for AI free will from systems that merely invite anthropomorphic projection. This vagueness undermines the paper's stated aim of providing a 'diagnostically useful' checklist, as different interpreters could reach opposite verdicts about the same system based on the same evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"The paper argues that AI systems can have free will if they satisfy three conditions: intentional agency (goal-directed action based on beliefs and desires), alternative possibilities (genuine choice between options), and causal control (high-level states make a difference to actions). Drawing on Dennett's intentional stance and a compatibilist framework, the author proposes that we should assess AI systems by asking whether they are best explained as intentional agents with these capacities. The paper addresses several objections (determinism, unpredictability, consciousness, simple algorithms) and concludes that free will in AI is a live possibility, though it does not claim that current systems definitively have it.","tokens_in":19135,"tokens_out":1753,"duration_ms":17614,"significance":"The paper provides a clear, pragmatically motivated framework that makes the question of free will in AI tractable and diagnostically useful. It synthesizes Dennett's ideas with the author's earlier work on free will, offering a level-of-description argument for how alternative possibilities can coexist with determinism at lower levels. The discussion of explainable AI as a driver for intentional-agent design is timely. The paper engages constructively with the existing literature and offers a novel, non-mysterious conception of free will that can be applied to both biological and artificial agents.","major_comments":[{"comment":"The paper relies on the requirement that the intentional stance be 'explanatorily clearly superior' to non-intentional alternatives in order to block over-ascription of free will to simple systems like thermostats and chess computers. However, no principled criteria for 'clearly superior' are provided; the paper merely appeals to practical feasibility and simplicity, which are context-dependent. Without an operationalized or at least more concretely specified test, the framework cannot reliably distinguish cases where the intentional stance is genuinely indispensable from cases where it is merely convenient. This vagueness undermines the diagnostic utility that the paper claims for its checklist (see also the skeptic note in the reviewer materials). The author should either offer more specific criteria or acknowledge this limitation more explicitly as a direction for future work.","section":"Section 5.4"}],"minor_comments":[{"comment":"Typo: 'refrain fom reading it' should be 'refrain from reading it'.","section":"Section 3"},{"comment":"The footnote says 'one might argue that if we need to refer to the original objective function to define the state change rule, we have not genuinely eliminated the choice-theoretic format.' This is an important point that could be developed; consider expanding it in the main text or at least clarifying its implications for the criterion.","section":"Section 5.4, footnote 14"},{"comment":"The analogy with corporate agents (citing List 2025) is interesting but somewhat brief; a short note on how the same three conditions apply to group agents could be included to strengthen the overall argument, or alternatively the reference could be flagged as supplementary reading.","section":"Section 6"},{"comment":"Some works cited in the text (e.g., Agüera y Arcas and Norvig 2024) are listed with the first author as 'y Arcas, B. A.'; this should be standardized to 'Agüera y Arcas, B.' or similar to match standard citation formats.","section":"Reference list"}],"recommendation":"minor_revision","confidential_remarks":"The paper is well-argued and makes a valuable contribution, but the central operational criterion ('explanatorily clearly superior') remains vague, which is a nontrivial concern for a framework that claims diagnostic utility. A minor revision could address this by adding a more explicit discussion of the criteria or by conceding that the framework currently functions primarily as a philosophical heuristic rather than a fully operational test. The paper relies heavily on the author's own previous work, but this is appropriate given that the current paper extends that work to the AI domain."},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Christian, you asked what I think of List's paper on free will in AI. The short version: it is a clean, well-written philosophical argument that applies his existing three-condition framework (intentional agency, alternative possibilities, causal control) to AI systems. That application is the main new contribution, and he handles the usual objections—determinism, unpredictability, consciousness, simple algorithms—with plausible, often deflationary responses. The paper does what it sets out to do: lay out criteria and argue that AI free will is not ruled out by standard objections.\n\nThe real soft spot is the one the stress-test note flags. The central claim is that a system has free will when we have good explanatory reasons to view it as meeting the three conditions, and specifically when the intentional stance is 'explanatorily clearly superior' to non-intentional alternatives (section 5.4). The paper concedes that simple optimizers and thermostats admit both descriptions and says we should not attribute free will to them because the non-intentional explanation is equally adequate. But it never gives a principled, operational test for when one explanation is clearly superior. 'Practical feasibility' and 'simpler descriptions' are context-dependent and potentially subjective. Without that, the framework cannot reliably distinguish a genuinely agential AI from a system that merely invites anthropomorphic projection. This is not a fatal flaw in a philosophical paper, but it does limit the claimed 'diagnostically useful' payoff.\n\nOtherwise, the argument is internally coherent and fairly grounded in the existing literature (though heavily self-cited). The paper is honest about its limitations—it does not claim to settle empirical questions about current AI. It is a piece of conceptual engineering, not an empirical result.\n\nWho is this for? Philosophers of mind and AI ethicists who want a pragmatic, Dennett-style account of what could count as artificial free will. It will be cited for its framing.\n\nMy recommendation: yes, send it to peer review. It is a competent, focused piece that makes a clear argument. The vagueness of the 'clearly superior' criterion is a genuine concern but one that reviewers can press during revision. A good referee can help sharpen that condition or make the paper more modest about its diagnostic ambitions.","headline":"Clear philosophical extension of List's free-will framework to AI, but the core diagnostic criterion is too vague to deliver on its promise of practical usefulness.","tokens_in":19648,"tokens_out":911,"would_cite":false,"duration_ms":10367,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"AI systems can qualify for free will if they meet three conditions: intentional agency, alternative possibilities, and causal control.","keywords":["artificial intelligence","free will","agency","autonomy","intentional stance","choice making","alternative possibilities","Daniel Dennett"],"falsifier":"If a state-of-the-art AI system (e.g., a large language model or autonomous robot) is subjected to the test of whether the non-intentional algorithmic explanation is equally or more parsimonious than the intentional-agent explanation for its behaviour across a wide range of cases, and the algorithmic explanation consistently wins, then the paper's claim that such systems can have free will would be falsified. Conversely, if explaining the system's behaviour in terms of beliefs, desires, and choices becomes indispensable for accurate prediction and understanding, the claim is supported.","tokens_in":20434,"feed_emoji":"🤖","tokens_out":2114,"duration_ms":20249,"temperature":0.7,"pith_summary":"This paper argues that the question of whether AI systems have free will should not be answered by looking for mysterious properties, indeterministic algorithms, or unpredictability. Instead, it proposes a pragmatic three-condition checklist inspired by Daniel Dennett: the system must be best explained as an intentional agent, must have alternative possibilities to choose from, and must have causal control over its actions. If a system meets all three, it counts as having free will in a practically relevant, non-mysterious sense. The paper suggests that some current or near-future AI systems may well satisfy these conditions, making free will in AI much less far-fetched than typically assumed. This matters because free will is often considered a precondition for moral responsibility, and the framework provides a diagnostic tool for ascribing agency and accountability to increasingly autonomous systems.","feed_headline":"AI systems can have free will under a pragmatic, three-condition test","feed_subtitle":"A philosopher argues that free will is not a myserious property but a practically assessable capacity for intentional agency.","key_machinery":"The three-condition checklist (intentional agency, alternative possibilities, causal control) derived from List (2019), combined with a level-relative notion of possibility that allows alternative possibilities at the macro-level of agency even when the micro-level is deterministic. The machinery also includes a realist reinterpretation of Dennett's intentional stance—treating the explanatory indispensability of agent-level descriptions as evidence, not mere interpretation.","core_discovery":"The paper establishes a set of necessary and sufficient conditions for free will that can be applied to AI systems: intentional agency (the system acts on beliefs and desires), alternative possibilities (the system faces genuine choices at the agential level), and causal control (the system's high-level states are difference-making causes of its actions). The central discovery is that free will is a system-level, explanatory property: if taking the 'intentional stance' toward an AI system is explanatorily superior to a purely mechanistic description, then that system qualifies as having free will. The paper defends this by arguing that determinism at the algorithmic level does not preclude alternative possibilities at the agentive level, and that consciousness is not required for free will. This shifts the debate from metaphysical speculation to empirical and practical assessment of AI agency.","pith_inferences":["The paper's pragmatic definition might be too weak for those who insist on libertarian free will or consciousness as necessary, but it offers a useful working definition for engineering and policy contexts where agency needs to be attributed.","A direct test of the framework would be to attempt building an AI system that explicitly satisfies the three conditions and then compare its explainability and accountability with a system that does not.","The framework implies that the 'hard problem' of free will (the metaphysical question) is less relevant than the practical question of when to treat a system as an agent, which aligns with current trends in AI ethics and explainability.","If future AI systems are designed to be maximally explainable in agentive terms, they would automatically be candidates for free will under this account, potentially accelerating the need for legal recognition of AI agency."],"forward_implications":["If an AI system is best explained as an intentional agent with alternative possibilities and causal control, then it has free will in a sense that underwrites moral responsibility (though not automatically moral agency).","The framework provides concrete criteria for designers and regulators to assess whether a given AI system should be treated as an autonomous agent rather than a mere tool.","Free will in AI becomes an empirically tractable question: we can test whether taking the intentional stance is genuinely superior to a purely algorithmic explanation for a given system.","The possibility of free will extends to other non-biological agents such as corporations and organized collectives, as they can also satisfy the three conditions.","AI systems with free will would qualify for a form of moral patiency and could in principle be held responsible for their actions, raising new ethical and legal considerations."],"supporting_citations":[{"why":"Supplies the three-condition definition of free will and the level-relative account of alternative possibilities that the paper applies to AI.","marker":"List 2019"},{"why":"Provides the intentional stance framework and the general approach of interpreting free will as a practically useful capacity rather than a mysterious property.","marker":"Dennett 1984, 1987, 2003"},{"why":"Supports the claim that it is empirically plausible for LLMs to have beliefs, a key step in establishing intentional agency.","marker":"Levinstein and Herrmann 2025"},{"why":"Directly discusses free will in AI using a decision-theoretic perspective, serving as a precursor and parallel argument.","marker":"Maier 2023"},{"why":"Provides evidence that LLMs develop world models, supporting the attribution of representational states (beliefs).","marker":"Gurnee and Tegmark 2023"},{"why":"Connects explainable AI to intentional explanations, reinforcing why AI systems might be designed to satisfy the three conditions.","marker":"Miller 2019"},{"why":"Discussed as a contrasting view that requires mindedness for free will, helping to define the paper's own position by contrast.","marker":"Floridi and Sanders 2004"}],"fun_headline_variants":["Three conditions determine if an AI has free will","Philosopher argues AI free will needs no consciousness","For AI, free will is a pragmatic property","To assess AI free will, ask if it makes real choices","AI free will requires agency not indeterminism"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The three conditions (intentional agency, alternative possibilities, causal control) are jointly sufficient for free will, and this definition is accepted as capturing the 'varieties of free will worth wanting'—if a reader rejects this definition (e.g., requiring consciousness or indeterminism), the argument for AI free will collapses.","fun_headline_variants_meta":{"raw":{"variants":["Three conditions determine if an AI has free will","Philosopher argues AI free will needs no consciousness","For AI, free will is a pragmatic property","To assess AI free will, ask if it makes real choices","AI free will requires agency not indeterminism"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"cost":0.000210868,"is_byok":false,"cost_details":{"upstream_inference_cost":0.000210868,"upstream_inference_prompt_cost":8.6548e-05,"upstream_inference_completions_cost":0.00012432},"total_tokens":1267,"prompt_tokens":823,"completion_tokens":444,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":256,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":368}},"tokens_in":823,"tokens_out":444,"cost_source":"openrouter_usage","duration_ms":6682,"image_pages":[],"temperature":1.0,"reasoning_tokens":368,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-09-15T08:19:31.717104+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"falsifier":"If a state-of-the-art AI system (e.g., a large language model or autonomous robot) is subjected to the test of whether the non-intentional algorithmic explanation is equally or more parsimonious than the intentional-agent explanation for its behaviour across a wide range of cases, and the algorithmic explanation consistently wins, then the paper's claim that such systems can have free will would be falsified. Conversely, if explaining the system's behaviour in terms of beliefs, desires, and choices becomes indispensable for accurate prediction and understanding, the claim is supported.","supporting_citations":[],"review_version":1}