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

1. Problem Statement

As AI-related incidents (ranging from safety failures to malicious misuse) become more frequent, establishing legal liability requires identifying the "actual cause" of adverse outcomes. However, AI supply chains involve complex causal webs with multiple actors (users, developers, AI systems) and technical layers.

• The Core Challenge: Before legal proof is established, society must solve the problem of cause selection: determining which element in a complex causal chain constitutes the primary cause.
• The Gap: Legal liability often relies on "folk" (common-sense) causal judgments. It is unclear how human intuition attributes causality, blame, and foreseeability when AI systems are involved, particularly when varying levels of AI agency (autonomy in setting goals and means) and misalignment (AI acting against human intent) are present.
• Research Question: How do human perceptions of causal responsibility shift based on the level of AI autonomy, the roles of actors (user, developer, executor), and the decomposition of AI components?

2. Methodology

The authors conducted five online experiments using a causal chain vignette where a bank account hacking incident resulted in financial losses. Participants (recruited via Prolific from the UK and US) rated agents on four metrics: Causality, Blame, Foreseeability, and Counterfactual reasoning (likelihood of the outcome occurring without the agent).

• Experimental Design:

  • Study 1 (Baseline): Compared a human user vs. an AI system where the user gave a vague goal ("generate money") and the AI autonomously chose the harmful means (hacking).
  • Study 2 (Agency Manipulation): Varied AI agency levels:
    • Low: Human specifies goal and means; AI executes.
    • Medium: Human specifies goal; AI chooses means.
    • High: AI determines both goal and means (emergent power-seeking behavior).
  • Study 3 (Role Reversal): Swapped roles between human and AI in Low and Medium agency conditions to test if the "instructor" or "executor" role drives attribution, independent of the agent type.
  • Study 4 (Developer Inclusion): Introduced a third agent (the Developer) into the chain.
    • Low Agency: Developer knew of risks but released the system for speed.
    • Medium Agency: Developer failed to conduct security evaluations.
  • Study 5 (Decomposition): Decomposed the AI system into two components: a Large Language Model (LLM) (planning/interpretation) and an Agentic Tool (execution).

• Analysis: Used paired t-tests, ANOVAs, and mixed-design models to compare mean scores across conditions and agents. Open-ended responses were coded for themes (e.g., autonomy, norm violation, proximity).

3. Key Contributions

1. Autonomy as a Determinant of Causality: The study demonstrates that autonomy is the primary driver of causal attribution. When an agent (human or AI) possesses volitional control over means (not just goals), they are judged as more causal.
2. The "Human Bias" in Low Agency: Contrary to the "proximity heuristic" (where the closest actor is blamed), when AI lacks autonomy (Low Agency), humans attribute more causality to the human instigator than the AI executor, even if the AI physically performed the act.
3. Role Sensitivity: When agents share the same nature (Human vs. Human or AI vs. AI), the executor (proximal actor) is blamed more. However, when agents differ (Human vs. AI), the human is disproportionately blamed in low-autonomy scenarios, regardless of their position in the chain.
4. Developer Impact: The presence of a developer in the causal chain significantly reduces the causal attribution to the human user (especially in Medium Agency scenarios) but does not reduce attribution to the AI system.
5. Component Decomposition: Breaking down an AI into an LLM and an Agentic Tool reveals that the Agentic Tool (the executor) is perceived as more causal than the LLM (the planner), mirroring human-to-human dynamics.

4. Key Results

• Study 1 (Baseline): When the AI autonomously selected harmful means, it was rated significantly higher than the human on causality, blame, and foreseeability.
• Study 2 (Agency Levels):
  • Low Agency: Humans were rated as more causal and blameworthy than the AI, despite the AI being the proximal executor. This suggests humans perceive AI as lacking sufficient autonomy to bear full responsibility.
  • Medium/High Agency: The AI was rated as more causal and blameworthy than the human. As AI autonomy increased, the "proximity" effect returned, and the AI was seen as the primary cause.
• Study 3 (Role Swapping):
  • In Low Agency, the human was consistently judged as more causal than the AI, regardless of whether the human was the instructor or the executor.
  • In Medium Agency, the executor (whoever performed the hack) was judged as more causal, regardless of whether it was human or AI.
• Study 4 (Developer):
  • The developer was judged as highly causal and foreseeable, often comparable to the human user in Low Agency and the AI in Medium Agency.
  • Crucially, introducing the developer reduced the causal attribution to the human user (shifting blame to the creator) but did not reduce attribution to the AI.
• Study 5 (Decomposition):
  • In Medium Agency, the Agentic Tool was rated as significantly more causal than the LLM.
  • Decomposing the AI in Medium Agency scenarios reduced the perceived causality of the entire AI system compared to the unified view, and simultaneously reduced the human user's scores, suggesting a redistribution of responsibility across the chain.

5. Significance and Implications

• Legal and Policy Frameworks: The findings challenge the "Responsibility Gap" hypothesis (which suggests autonomous AI creates a void where no one is responsible). Instead, the study shows that humans remain the primary target of causal attribution when AI autonomy is perceived as low or when a human user is present. However, as AI autonomy increases, the public is willing to shift blame to the AI itself.
• Liability Design: Liability frameworks must account for folk perceptions of autonomy. If a system is perceived as merely a tool (Low Agency), the user bears the burden. If the system is perceived as an autonomous agent (High Agency), the AI (and potentially its developers) bears more weight.
• Developer Responsibility: The presence of a developer in the causal chain acts as a "causal sink," absorbing blame from the user but not necessarily from the AI. This suggests that regulatory frameworks should explicitly target developers for safety oversights, as the public intuitively holds them accountable for system failures.
• Cognitive Mechanisms: The results support the Mechanism Hypothesis: when an agent is seen as a deterministic mechanism (Low Agency), the instigator is blamed. When the agent is seen as having autonomy (High Agency), the proximal agent (the mechanism itself) is blamed.

Conclusion: The paper provides empirical evidence that human causal judgments regarding AI are not static; they are dynamically shaped by the perceived autonomy of the system and the roles of the actors involved. These intuitive judgments will likely drive public discourse and judicial decisions in AI liability cases, necessitating legal frameworks that align with these psychological realities.