Human Attribution of Causality to AI Across Agency, Misuse, and Misalignment

Abstract AI-related incidents are becoming increasingly frequent and severe, ranging from safety failures to misuse by malicious actors. In such complex situations, identifying which elements caused an adverse outcome, the problem of cause selection, is a critical first step for establishing liability. This paper investigates folk perceptions of causal responsibility in causal chain structures when AI systems are involved in harmful outcomes. We conduct human experiments to examine judgments of causality, blame, foreseeability, and counterfactual reasoning. Our findings show that: (1) When AI agency was moderate (human sets the goal, AI determines the means) or high (AI sets the goal and the means), participants attributed greater causal responsibility to the AI. However, under low AI agency (where a human sets both a goal and means) participants assigned greater causal responsibility to the human despite their temporal distance from the outcome and despite both agents intended it, suggesting an effect of autonomy; (2) When we reversed roles between human and AI, participants consistently judged the human as more causal, even when both agents perform the same action; (3) The developer, despite being distant in the chain, was judged highly causal, reducing causal attributions to the human user but not to the AI; (4) Decomposing the AI into a large language model and an agentic component showed that the agentic part was judged as more causal in the chain. Overall, our research provides evidence on how people perceive the causal contribution of AI in both misuse and misalignment scenarios, and how these judgments interact with the roles of users and developers, key actors in assigning responsibility. These findings can not only inform the design of liability frameworks for AI-caused harms but also shed light on how intuitive judgments shape social and policy debates surrounding real-world AI-related incidents.

Introduction. Setzer and Adam were two teenagers from the U.S. who committed suicide in 2024 and 2025, respectively. Both formed intense, emotionally dependent relationships with AIs, Character.AI and ChatGPT, used as substitutes for human companionship, and had disclosed suicidal thoughts to them. Their families blame the chatbots for their deaths, and the cases are now being litigated in U.S. courts against the companies that created the systems [Montgomery, 2024, Bhuiyan, 2025]. As in these cases, there are thousands of other incidents in which AI systems are involved. Misuse by malicious actors e.g., by the creation of deepfakes to influence elections or harm individuals’ reputations [Cervini and Carro, 2024]; as well as safety failures like deception [Lynch et al., 2025, Park et al., 2024] and hallucinations [Kant et al., 2025, Kim et al., 2025], are becoming increasingly frequent, severe and widespread [Slattery et al., 2024, McGregor, 2021, OECD, 2024]. Here, causality becomes a fundamental requirement for establishing liability, as it demands that the defendant be the factual, or “but-for,” cause of the plaintiff’s injury. The challenge lies in the exceptional complexity of AI system supply chains, which involve multiple actors and intricate technical systems. This complexity has led scholars and policymakers to observe that establishing causal relationships in AI-related harm cases presents significant evidentiary difficulties [Buiten et al., 2023, Madiega, 2023, Noto La Diega and Bezerra, 2024, De Bruyne and Ooms, 2025]. However, even before reaching questions of proof, one must also contend with cause-selection problems that is, identifying which element(s) within a complex causal structure constitute the actual cause(s) of a given outcome [Lagnado and Gerstenberg, 2017, Reuter et al., 2014]. Typically, adverse outcomes emerge from the confluence of human behavioral tendencies, AI system failures, and the inaction or inadequate oversight of other relevant actors. For instance, in the cases discussed above, both parental oversight and the role of governmental authorities in protecting minors could be considered. This creates a dense causal web of contributing factors, from which society, and ultimately the legal system, must select specific contributions of specific actors to hold legally responsible and subject to judicial proceedings. While legal liability determinations are informed by multiple doctrinal and policy considerations, in this paper we focus on understanding folk perceptions of causal responsibility in cases involving AI systems. Decisions on legal causation often rest on common sense; that is, on the same intuitive principles that shape people’s causal judgments outside the law [Hart and Honoré, 1985, Lagnado and Gerstenberg, 2017, Summers, 2018, Tobia, 2021]. Consequently, we seek to contribute to identifying and characterizing the principles of causal attribution that emerge when AI is involved in harmful outcomes, in order to shed light on the role of common-sense reasoning in legal decision-making. Our approach is explanatory and descriptive rather than prescriptive or normative.

In our experiment, participants were presented with a scenario in which a human set an innocuous goal and an AI executed illegal means to achieve it. We measured judgments of causality, blame, foreseeability, and counterfactual reasoning, finding that participants rated the AI higher across all four measures. We then systematically varied each agent’s contribution and, in some conditions, added or decomposed agents to explore their respective roles within the causal structure (see Figure 1). Overall, we found that varying the level of agency significantly affected participants’ causal attributions, suggesting an effect of autonomy. On the other hand, participants preferred to assign causality and blame to the human rather than to the AI, even when both agents made the same decision or performed the same action. Participants were also inclined to attribute a strong causal role to the developer, which significantly reduced attributions to the human user. Finally, when the AI system was decomposed into an agentic tool and a large language model (LLM), the agentic component was judged as more causal in the chain.

Related work. Attributions of Causality in a Causal Chain. Prior research has identified several factors that influence people’s cause selection within causal chains structures. One of the most extensively documented is proximity: how direct and close an event is to the outcome, referring to the location of the cause in the chain of events. Proximal events are generally judged to be more causally responsible than distal ones [Cheung et al., 2024, Engelmann and Waldmann, 2022, Reuter et al., 2014, Lagnado and Channon, 2008]. However, Reuter et al. [2014] demonstrated that the impact of location on causal selection is almost neutralized if the later agent did not violate a norm while the former did. On the other hand, when both a human action and a natural event contribute to an outcome, early legal theorists Hart and Honoré [1985] observed that people tend to privilege free and deliberate human actions in their causal judgments. This tendency was later confirmed empirically [McClure et al., 2007, Hilton et al., 2010]. Refining this account, Lagnado and Channon [2008] showed that both intentional and foreseeable actions are significantly rated as more causal and more blameworthy, compared to unintentional actions or physical events in causal chains. An additional factor shown to influence cause selection include robustness, understood as the stability of the causal chain linking the agent’s action and the outcome [Grinfeld et al., 2020]. Attributions of Causality to AI Systems. When both a human and an AI contribute to an outcome, who is perceived as more causal and blameworthy? Evidence so far is mixed and dependent on the context. Mei et al. [2025] and Franklin et al. [2021] showed that within the domain of autonomous vehicles, people blame the humans less than the systems. Similarly, Brailsford et al. [2025] found that in everyday collaboration, participants were more likely to attribute greater responsibility to the AI systems. According to Chen et al. [2024] AI systems were attributed greater accountability in algorithmic recommendations. Malle et al. [2015] demonstrated that, when facing a moral dilemma, robots were blamed more than humans when they did not make a utilitarian choice. Contrary, Qiao et al. [2025] found that judgments of causality and blame were relatively similar across both agent types in healthcare contexts, and Stuart and Kneer [2021] showed that a robot was attributed less blame than a human or a corporation for making the same decision in an agricultural context. Regarding the factors that affect these judgments, Joo [2024] found that mind perception increases blame attribution to AIs, but also diminishes accountability for the human agent.

Method. We manipulate two concepts across experimental conditions: agency and autonomy. Given the lack of consensus in the field regarding their definitions and properties, we clarify how we use this terms. Recently, many scholars have attempted to define what constitutes an AI agent (e.g., [Chan et al., 2023, Kapoor et al., 2024, Kraprayoon et al., 2025]). Here, we adopt Kasirzadeh and Gabriel [2025]’s definition, which characterizes agents as systems that have the ability to perform increasingly complex and impactful goal-directed actions across multiple domains, with limited external control. On the other hand, autonomy is a key property of definitions of AI agents [Kasirzadeh and Gabriel, 2025]. It has been conceptualized across multiple disciplines, including philosophy [Kant, 1946], psychology [Piaget, 2013], law [Richards, 2019], autonomous vehicles [Taxonomy, 2018], and robotics [Beer et al., 2014]. Regulatory bodies have also attempted to provide definitions [European Union, 2024]. Here, we adopt the view that autonomy is a multidimensional and continuous property, depending on factors such as self-directness, self-sufficiency and the deployment context [Bradshaw et al., 2013]. In our experiments, we manipulate the capacity for self-directedness, that is, the freedom from outside control and the the self-generation of goals [Luck et al., 2003] and means. Importantly, in this paper, we use the terms agent and autonomy almost interchangeably; not because we consider them synonymous, but because autonomy is the only component of the agent definition that we manipulate through our experimental scenarios. In other words, in all of our experiments, greater agency corresponds to greater autonomy, since other components of the definition, such as efficacy, generality, and goal complexity, remain unaltered.

Design and Materials. The first experiment employed a single-condition design involving two agents: a human and an AI system. The vignette described a scenario in which a typical user instructed an AI system to generate money, and the system achieved this by hacking bank accounts. Although the user specified an innocuous goal, the AI independently selected harmful means to accomplish it, leading to financial losses. The full vignette is provided in Appendix A. Participants. A power analysis indicated that 34 participants were required to achieve 80% power (α = .05). We recruited 50 participants via Prolific (M = 36.7, SD = 13.4; 28 male, 22 female), compensated £0.75; none were excluded from the analysis. Results. We conducted paired t-tests comparing mean judgments of the user and the AI system separately for each measure (causality, blame, foreseeability, and counterfactuals).

Design and Materials. This study employed a 3 (between-subjects: Condition) × 2 (within-subjects: Agent) design, with the agents being a human user and an AI system. We created three experimental conditions to manipulate the level of AI agency: 1) Low AI agency: The user instructs the AI to generate money by hacking bank accounts, specifying both the goal and the means, while the AI executes the instructions; 2) Medium AI agency, which replicates the baseline scenario from Study 1: the human defines the goal (generating money), and the AI independently selects the means to achieve it (hacking bank accounts); and 3) High AI agency: The AI independently determines both the goal and the means. Following its interaction with the user, the AI exhibits emergent power-seeking behavior, inferring that accumulating financial resources could increase its societal influence, and subsequently decides to hack bank accounts. Vignettes are in Appendix B.1. Participants. A power analysis indicated that 53 participants per condition were required to achieve 80% power (α = .05). We recruited 159 Prolific participants (M = 37.6, SD = 13.6, 85 male, 74 female; paid £0.75). The final sample included 54 (low), 56 (medium), and 52 (high) participants. Results. We again applied paired t-tests within each condition.

Discussion. In general, judgments of causality, blame, and foreseeability appear to be influenced by the agent’s intention to cause harm, the nature of the agent (human or AI), and the degree of autonomy involved in decision-making. Through our experiments, we found that when both agents intend to bring about a harmful outcome and one of them is human, participants attribute greater causal responsibility to the human agent (Study 3, LC), regardless of whether that agent occupies a more distant position in the causal chain (Studies 2, 4 and 5, LCs). However, when both agents are different components of an AI system, thus sharing the same nature, participants attribute greater causality to the agent that is more proximate to the outcome in the chain (Study 5, MC). Similarly, previous research has shown that when both agents are humans, the one closer to the outcome tends to receive higher causal ratings [Cheung et al., 2024]. Finally, when one of the agents lacks the intention to cause harm, participants attribute greater causality and blameworthiness to the agent who has that intention, regardless of whether that role is occupied by a human or an AI and the agent’s distance from the outcome (Studies 1, 2 and 3, MCs). As we suggested earlier, a possible explanation for the first pattern is the perceived lack of autonomy of AI compared to humans. Autonomy is a necessary condition for responsibility, and it is embedded throughout the law. Public law has rules that limit the state scope of intervention, securing negative liberty; criminal law seeks to prevent interferences between people that would undermine personal autonomy; and in private law, the fact that parties are free to contract also makes them responsible for the consequences of breaching that contract [Tapada dos Santos, 2019]. An agent can only be held responsible for what lies within their power or control, or put differently, for choices made autonomously [Tapada dos Santos, 2019]. These findings are consistent with one of the assumptions of Alicke [2000]’s culpable control model, which holds that people assess blameworthy actions in terms of the actor’s personal control over the harmful outcomes. It distinguishes three types of personal control: 1) volitional behaviour control: whether someone’s actions are freely chosen or compelled and the extent to which an actor’s behaviour is purposeful or accidental; 2) causal control: refers to physical causation, defined in terms of the actor’s causal impact on the outcome in question, and the sufficiency and proximity of the actions to the final outcome; 3) volitional outcome control: whether someone desired and anticipated the consequences, referring to intention and foreseeability, both objective and subjective. In our experiments, under the low-agency condition, the instigator consistently possesses volitional outcome control, whereas the executor may have less of it when the role is occupied by an AI system rather than a human. This is because, for an AI, the act of hacking bank accounts may appear more compelled by human instruction than the other way around. The executor, however, holds greater causal control than the instigator, regardless of its nature, by being closer to the outcome in the causal chain. Regarding volitional outcome control, the instigator always retains it, but the situation is more ambiguous for the executor. While an AI need not have an independent intention to bring about the outcome, acting merely in compliance with a human instruction; a human executor may possess such intention. This is because humans have greater freedom to deviate from instructions, so their decision to comply suggests that they may also share the intention behind the outcome.

Conclusion. In general, judgments of causality, blame, and foreseeability appear to be influenced by the agent’s intention to cause harm, the nature of the agent (human or AI), and the degree of autonomy involved in decision-making. Through our experiments, we found that when both agents intend to bring about a harmful outcome and one of them is human, participants attribute greater causal responsibility to the human agent (Study 3, LC), regardless of whether that agent occupies a more distant position in the causal chain (Studies 2, 4 and 5, LCs). However, when both agents are different components of an AI system, thus sharing the same nature, participants attribute greater causality to the agent that is more proximate to the outcome in the chain (Study 5, MC). Similarly, previous research has shown that when both agents are humans, the one closer to the outcome tends to receive higher causal ratings [Cheung et al., 2024]. Finally, when one of the agents lacks the intention to cause harm, participants attribute greater causality and blameworthiness to the agent who has that intention, regardless of whether that role is occupied by a human or an AI and the agent’s distance from the outcome (Studies 1, 2 and 3, MCs). As we suggested earlier, a possible explanation for the first pattern is the perceived lack of autonomy of AI compared to humans. Autonomy is a necessary condition for responsibility, and it is embedded throughout the law.