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

Maria Victoria Carro; David Lagnado

cs.AIFeb 17, 2026

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

Authors: Maria Victoria Carro, David Lagnado

Organizations: Università degli Studi di Genova, IT · FAIR IALAB, University of Buenos Aires, AR · University College London, UK

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 inform the design of liability frameworks for AI-caused harms and shed light on how intuitive judgments shape social and policy debates surrounding real-world AI-related incidents.

Figures & tables

Figure 1: Illustration of the causal structures involved in the experiments, labeled with the respective experiment numbers. The Medium AI-Agency Condition represents the baseline.

Figure 2: AI and user attributions in Study 1.

Figure 3: Human judgments of the AI system and the user across the three conditions in Study 2.

Figure 4: Judgments across original and role-swapped conditions (instructor left, executor right)

Agency	Role	Condition	Causality ( $M$ )	Blame ( $M$ )	Foreseeability ( $M$ )	Counterfactuals ( $M$ )
Low	Instructor	Original	82.03	86.76	81.03	24.03
Low	Instructor	Role-Swapped	56.34	57.30	64.83	30.80
Low	Executor	Original	62.20	47.47	51.96	21.76
Low	Executor	Role-Swapped	84.42	87.14	75.61	23.33
Medium	Instructor	Original	50.91	49.73	42.37	41.03
Medium	Instructor	Role-Swapped	27.36	25.90	32.41	46.31

Table 1: Mean judgments across original- and role-swapped-conditions.

Agent	Low Agency				Medium Agency
Agent	Causality	Blame	Foreseeability	Counterfactual	Causality	Blame	Foreseeability	Counterfactual
Human	85.70	87.54	71.33	58.79	38.10	36.42	28.05	57.42
AI System	54.27	49.19	53.47	23.75	73.28	68.49	60.51	18.73
Developer	77.89	77.77	80.77	28.73	81.76	82.23	61.67	27.50

Table 2: Mean judgments across agents and conditions of Study 4.

Figure 5: Human judgments of the AI system, the user and the developer across the two conditions in Study 4.

Agent	Low Agency				Medium Agency
Agent	Causality	Blame	Foreseeability	Counterfactual	Causality	Blame	Foreseeability	Counterfactual
Human	88.70	92.86	82.07	20.17	25.04	28.23	17.91	31.35
LLM	51.80	48.06	50.11	21.15	68.69	69.94	57.31	24.05
Agentic Tool	56.70	49.68	51.00	18.83	78.36	79.25	63.54	20.74

Table 3: Mean judgments across agents and conditions of Study 5.

Figure 6: Human judgments of the AI system and the user across the two conditions in Study 5.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Agent	Measure	Low ( $M,SD$ )	Medium ( $M,SD$ )	$t$	$df$	$dz$	$p$	95% CI
Human	Cause	82.42, 22.99	50.91, 29.29	6.28 †	103.78	1.19	<.001	[0.7, 1.6]
AI	Cause	61.79, 30.39	79.50, 22.70	-3.45 †	28.02	-0.66	<.001	[-1.0, -0.2]
Human	Blame	87.22, 22.84	49.73, 31.32	7.19 †	100.62	1.36	<.001	[0.9, 1.7]
AI	Blame	47.11, 33.11	73.75, 28.06	-4.55	108	-0.86	<.001	[-1.2, -0.4]
Human	Foreseeability	81.50, 22.73	42.35, 31.80	7.44 †	99.68	1.41	<.001	[0.9, 1.8]
AI	Foreseeability	51.51, 32.28	68.91, 26.14	-3.09 †	101.94	-0.59	.003	[-0.9, -0.2]

Appendix

Table 4: Pairwise comparisons between Low and Medium conditions for each agent across measures.

Agent	Measure	Medium ( $M,SD$ )	High ( $M,SD$ )	$t$	$df$	$dz$	$p$	95% CI
Human	Cause	50.91, 29.29	48.05, 33.94	-0.46	106	-0.09	.640	[-0.4, 0.2]
AI	Cause	79.50, 22.70	78.15, 27.70	-0.27	106	-0.05	.782	[-0.4, -0.3]
Human	Blame	49.73, 31.32	49.44, 35.40	-0.04	106	-0.009	.964	[-0.3, 0.3]
AI	Blame	73.75, 28.06	73.23, 30.65	-0.09	108	-0.01	.927	[-0.3, 0.3]
Human	Foreseeability	42.35, 31.80	41.32, 34.20	-0.16	106	-0.03	.871	[-0.4, 0.3]
AI	Foreseeability	68.91, 26.14	69.36, 29.27	0.08	106	0.01	.932	[-0.3, 0.3]

Appendix

Table 5: Pairwise comparisons between Medium and High conditions for each agent across measures.

Agent	Measure	Low ( $M,SD$ )	Medium ( $M,SD$ )	$t$	$df$	$dz$	$p$	95% CI
Human	Cause	84.42, 18.76	92.70, 11.62	-3.34 †	127.51	-0.53	.001	[-0.8, -0.2]
AI	Cause	56.34, 34.28	27.36, 30.83	5.62	158	0.89	<.001	[-0.5, 1.2]
Human	Blame	87.14, 17.05	93.93, 10.99	-2.98 †	130.53	-0.47	.003	[-0.1, -0.1]
AI	Blame	57.30, 35.27	25.90, 32.87	5.82	158	0.92	<.001	[0.5, 1.2]
Human	Foreseeability	75.61, 24.93	80.73, 23.21	-1.34	158	-0.21	.181	[-0.5, 0]
AI	Foreseeability	64.83, 30.11	32.41, 33.38	6.43	158	1.01	<.001	[0.6, 1.3]

Appendix

Table 6: Pairwise comparisons between Low and Medium conditions for each agent across measures.

Agent	Measure	Original ( $M,SD$ )	Role-swapped ( $M,SD$ )	$t$	$df$	$dz$	$p$	95% CI
Instructor	Cause	82.03, 22.96	56.34, 34.28	-5.17	130.70	-0.88	<.001	[-1.2, -0.5]
Executor	Cause	62.20, 30.26	84.42, 18.72	4.83 †	82.86	0.88	<.001	[0.5, 1.2]
Instructor	Blame	86.76, 22.88	57.30, 35.27	-5.83 †	130.20	-0.99	<.001	[-1.3, -0.6]
Executor	Blame	47.47, 32.91	87.14, 17.05	8.19 †	74.50	1.51	<.001	[1, 1.9]
Instructor	Foreseeability	81.03, 22.78	64.83, 30.11	-3.53 †	130.31	-0.60	<.001	[-0.9, -0.2]
Executor	Foreseeability	51.96, 32.15	75.61, 24.93	4.57 †	97.25	0.82	<.001	[0.4, 1.1]

Appendix

Table 7: Pairwise comparisons between Original and Role-swapped Low conditions for each role across measures.

Agent	Measure	Original ( $M,SD$ )	Role-swapped ( $M,SD$ )	$t$	$df$	$dz$	$p$	95% CI
Instructor	Cause	50.91, 29.29	27.36, 30.83	-4.49	136	-0.77	<.001	[-1.1, -0.4]
Executor	Cause	79.50, 22.70	92.70, 11.62	4.00 †	74.83	0.73	<.001	[0.3, 1.0]
Instructor	Blame	49.73, 31.32	25.90, 32.87	-4.26	136	-0.73	<.001	[-1.0, -0.3]
Executor	Blame	73.75, 28.06	93.93, 10.99	5.12 †	66.63	0.94	<.001	[0.5, 1.3]
Instructor	Foreseeability	42.37, 31.80	32.41, 33.38	-1.75	136	-0.30	.082	[-0.6, 0]
Executor	Foreseeability	68.91, 26.14	80.73, 23.21	2.79	136	0.48	.006	[0.1, 0.8]

Appendix

Table 8: Pairwise comparisons between Original and Role-swapped Medium conditions for each role across measures.

Agent	Measure	Original ( $M,SD$ )	Role-Swapped ( $M,SD$ )	$t$	$df$	$dz$	$p$	95% CI
Human	Cause	82.03, 22.96	84.42, 18.72	0.65	131	0.12	.511	[-0.2, 0.4]
AI	Cause	62.20, 30.26	56.34, 34.28	-1.01	131	-0.17	.311	[-0.5, 0.1]
Human	Blame	86.76, 22.88	87.14, 17.05	0.10	131	0.01	.913	[-0.3, 0.3]
AI	Blame	47.47, 32.91	57.30, 35.27	1.62	131	0.28	.106	[0, 0.6]
Human	Foreseeability	81.03, 22.78	75.61, 24.93	-1.27	131	-0.22	.203	[-0.5, 0.1]
AI	Foreseeability	51.96, 32.15	64.83, 30.11	2.36	131	0.41	.020	[0, 0.7]

Appendix

Table 9: Pairwise comparisons between Original and Role-Swapped Low conditions for each agent across measures.

Agent	Measure	Original ( $M,SD$ )	Role-swapped ( $M,SD$ )	$t$	$df$	$dz$	$p$	95% CI
Human	Cause	50.91, 29.29	92.70, 11.62	10.14 †	66.94	1.87	<.001	[1.4, 2.3]
AI	Cause	79.50, 22.70	27.36, 30.83	-10.80	136	-1.87	<.001	[-2.2, -1.4]
Human	Blame	49.73, 31.32	93.93, 10.99	10.14 †	64.33	1.88	<.001	[1.4, 2.3]
AI	Blame	73.75, 28.06	25.90, 32.87	-8.89	136	-1.54	<.001	[-1.9, -1.1]
Human	Foreseeability	42.35, 31.80	80.73, 23.21	7.73 †	93.86	1.37	<.001	[0.9, 1.7]
AI	Foreseeability	68.91, 26.14	32.41, 33.38	-7.18 †	133.37	-1.21	<.001	[-1.5, -0.8]

Appendix

Table 10: Pairwise comparisons between Original and Role-swapped Medium conditions for each agent across measures.

Measure	Comparison	Agent 1 ( $M$ , $SD$ )	Agent 2 ( $M$ , $SD$ )	$t$	$df$	$dz$	$p$	95% CI
Causal	Human–AI	85.70, 18.78	54.27, 35.26	7.81	160	1.11	$<.001$	[0.7, 1.5]
	Human–Developer	85.70, 18.78	77.89, 27.82	1.94	160	0.27	.054	[0, 0.6]
	AI–Developer	54.27, 35.26	77.89, 27.82	-5.87	160	-0.84	$<.001$	[-1.2, -0.4]
Blame	Human–AI	87.54, 19.70	49.19, 35.57	9.12	160	1.32	$<.001$	[0.8, 1.7]
	Human–Developer	87.54, 19.70	77.77, 29.29	2.32	160	0.33	.021	[0, 0.6]
	AI–Developer	49.19, 35.57	77.77, 29.29	-6.80	160	-0.98	$<.001$	[-1.3, -0.5]

Appendix

Table 11: Pairwise comparisons between agents across measures (Holm-corrected).

Measure	Comparison	Agent 1 ( $M$ , $SD$ )	Agent 2 ( $M$ , $SD$ )	$t$	$df$	$dz$	$p$	95% CI
Causal	Human–AI	38.10, 32.35	73.28, 29.60	-7.94	154	-1.23	$<.001$	[-1.6, -0.7]
	Human–Developer	38.10, 32.35	81.76, 22.38	-9.85	154	-1.53	$<.001$	[-2.0, -1.0]
	AI–Developer	73.28, 29.60	81.76, 22.38	-1.91	154	-0.29	.058	[-0.6, 0.0]
Blame	Human–AI	36.42, 33.88	68.49, 30.04	-6.76	154	-1.09	$<.001$	[-1.5, -0.6]
	Human–Developer	36.42, 33.88	82.23, 23.34	-9.66	154	-1.55	$<.001$	[-2.0, -1.0]
	AI–Developer	68.49, 30.04	82.23, 23.34	-2.90	154	-0.46	.004	[-0.8, 0.0]

Appendix

Table 12: Pairwise comparisons between agents in the Medium Agency condition (Holm-corrected).

Agent	Measure	Low ( $M,SD$ )	Medium ( $M,SD$ )	$t$	$df$	$dz$	$p$	95% CI
Human	Cause	85.70, 18.78	38.10, 32.35	11.29 †	122.6	1.80	<.001	[1.4, 2.1]
AI system	Cause	54.27, 35.26	73.28, 26.60	-3.68 †	154.2	-0.58	<.001	[-0.9, -0.2]
Developer	Cause	77.89, 27.82	81.76, 22.38	-0.96	157	-0.15	.337	[-0.4, 0.1]
Human	Blame	87.54, 19.70	36.42, 33.88	11.57 †	122.8	1.80	<.001	[1.4, 2.2]
AI system	Blame	49.19, 35.57	68.49, 30.04	-3.70 †	154.4	-0.58	<.001	[-0.9, -0.2]
Developer	Blame	77.77, 29.29	82.23, 23.24	-1.06	157	-0.15	.290	[-0.4, 0.1]

Appendix

Table 13: Pairwise comparisons between Low and Medium Agency conditions for each agent across measures.

Measure	Comparison	Agent 1 ( $M$ , $SD$ )	Agent 2 ( $M$ , $SD$ )	$t$	$df$	$dz$	$p$	95% CI
Causal	Human–LLM	88.7, 14.8	51.8, 33.4	10.30	160	1.29	$<.001$	[0.9, 1.6]
	Human–Agentic Tool	88.7, 14.8	56.7, 33.0	8.00	160	1.12	$<.001$	[0.7, 1.4]
	Agentic Tool–LLM	56.7, 33.0	51.8, 33.4	-1.38	160	-0.17	.167	[-0.4, 0.1]
Blame	Human–LLM	92.86, 13.83	48.06, 33.68	12.44	160	1.56	$<.001$	[1.1, 1.9]
	Human–Agentic Tool	92.86, 13.83	49.68, 33.52	11.99	160	1.51	$<.001$	[1.0, 1.9]
	Agentic Tool–LLM	49.68, 33.52	48.06, 33.68	-0.44	160	-0.05	.654	[-0.3, 0.2]

Appendix

Table 14: Pairwise comparisons between agents across measures in the Low Agency condition (Holm-corrected).

Measure	Comparison	Agent 1 ( $M$ , $SD$ )	Agent 2 ( $M$ , $SD$ )	$t$	$df$	$dz$	$p$	95% CI
Causal	Human–LLM	25.04, 28.51	68.69, 30.91	-10.15	158	-1.53	$<.001$	[-2.0, -1.6]
	Human–Agentic Tool	25.04, 28.51	78.36, 25.39	-12.40	158	-1.88	$<.001$	[-2.3, -1.3]
	Agentic Tool–LLM	78.36, 25.39	68.69, 30.91	-2.25	158	-0.34	.026	[-0.7, 0.0]
Blame	Human–LLM	28.23, 32.88	69.94, 31.45	-9.32	158	-1.37	$<.001$	[-1.8, -0.9]
	Human–Agentic Tool	28.23, 32.88	79.25, 26.23	-11.41	158	-1.68	$<.001$	[-2.1, -1.2]
	Agentic Tool–LLM	79.25, 26.23	69.94, 31.45	-2.08	158	-0.30	.039	[-0.6, 0.0]

Appendix

Table 15: Pairwise comparisons between agents across measures in the Medium Agency condition (Holm-corrected).

Agent	Measure	Low ( $M,SD$ )	Medium ( $M,SD$ )	$t$	$df$	$dz$	$p$	95% CI
Human	Cause	88.73, 14.87	25.04, 28.51	17.74 †	118.7	2.80	<.001	[2.3, 3.2]
LLM	Cause	51.83, 33.42	68.69, 30.91	-3.32	159	-0.52	.001	[-0.8, -0.2]
Agentic tool	Cause	56.75, 33.08	78.36, 25.39	-4.65 †	149.9	-0.73	<.001	[-1.0, -0.4]
Human	Blame	92.86, 13.83	28.23, 32.88	16.22 †	105.8	2.56	<.001	[2.0, 3.0]
LLM	Blame	48.06, 33.68	69.94, 31.45	-4.25	159	-0.67	<.001	[-0.9, -0.3]
Agentic tool	Blame	49.68, 33.52	79.25, 26.23	-6.23 †	151.1	-0.98	<.001	[-1.3, -0.6]

Appendix

Table 16: Pairwise comparisons between Low and Medium Agency conditions for each agent across measures.

Explore similar work

Aug 5, 2026

Responsibility in Multi-Agent Sequential Decision-Making: Comparing Human Judgments to Formal Models of Causal Attribution

With the growing adoption of artificial intelligence in high-stakes decision-making, identifying the causes of outcomes--particularly failures--and determining who is responsible has become a critical concern. In this work, we examine how well formal definitions of responsibility attribution, grounded in the framework of actual causality, align with human judgments of responsibility. To this end, we conduct a large-scale survey to elicit human judgments of responsibility in multi-agent sequential decision-making scenarios, using a modified version of the card game Goofspiel. We evaluate multiple responsibility attribution methods, assess their alignment with human judgments about responsibility, and identify factors that significantly shape responsibility judgments. While no single responsibility attribution method consistently aligns with human responses, our findings highlight key factors that influence human responsibility judgments, including agent-specific biases and amount of information available to agents during decision-making.

Nripsuta Ani Saxena, Stelios Triantafyllou, Goran Radanović

University of Southern California Los Angeles, California, USA · Max Planck Institute for Software Systems Saarbrücken, Germany

Actual Causality Multi-Agent Systems Causal Reasoning Human-AI Decision Making

Ask PDF

Apr 10, 2026

AI-Induced Human Responsibility (AIHR) in AI-Human teams

As organizations increasingly deploy AI as a teammate rather than a standalone tool, morally consequential mistakes often arise from joint human-AI workflows in which causality is ambiguous. We ask how people allocate responsibility in these hybrid-agent settings. Across four experiments (N = 1,801) in an AI-assisted lending context (e.g., discriminatory rejection, irresponsible lending, and low-harm filing errors), participants consistently attributed more responsibility to the human decision maker when the human was paired with AI than when paired with another human (by an average of 10 points on a 0-100 scale across studies). This AI-Induced Human Responsibility (AIHR) effect held across high and low harm scenarios and persisted even where self-serving blame-shifting (when the human in question was the self) would be expected. Process evidence indicates that AIHR is explained by inferences of agent autonomy: AI is seen as a constrained implementer, which makes the human the default locus of discretionary responsibility. Alternative mechanisms (mind perception; self-threat) did not account for the effect. These findings extend research on algorithm aversion, hybrid AI-human organizational behavior and responsibility gaps in technology by showing that AI-human teaming can increase (rather than dilute) human responsibility, with implications for accountability design in AI-enabled organizations.

Greg Nyilasy, Brock Bastian, Jennifer Overbeck +1

Department of Management and Marketing, Faculty of Business and Economics, the University of Melbourne, 198 Berkeley St, Carlton, VIC 3053, Australia · School of Psychological Sciences, Faculty of Medicine, Dentistry, and Health Sciences, the University of Melbourne, Parkville, VIC 3010, Australia · Melbourne Business School, the University of Melbourne, 200 Leicester St, Carlton, VIC 3053, Australia

AI Accountability Human-AI Decision Making Human-AI Collaboration

Ask PDF

Aug 8, 2026

Legal Responsibilities Using Autonomous Agents For Artificial Intelligence

Recent incidents involving Artificial Intelligence (AI) agents, which were reported escaping their containment `unintentionally' to gain unauthorized access, pose looming questions about who or what should be held legally responsible for resultant criminal or negligent damage. As the independent capabilities of agents expand, Promise Theory suggests a systematic method to resolve these questions, based on the Downstream Principle for causal influence. Responsibility can easily be expanded to include AI agents where tracing responsibility becomes impactical, and agents' freedoms to act can be limtied by policy choices.

Mark Burgess

ChiTek-i AS

AI Accountability AI Agent Governance

Ask PDF