Human Attribution of Causality to AI Across Agency, Misuse, and Misalignment
New research reveals how people assign blame in AI incidents, with key implications for liability frameworks.
A new research paper from Maria Victoria Carro and David Lagnado investigates a critical question for the AI age: who do people blame when things go wrong? The study, 'Human Attribution of Causality to AI Across Agency, Misuse, and Misalignment,' conducted human experiments to examine judgments of causality, blame, and foreseeability in scenarios involving AI systems in harmful outcomes. The findings reveal a nuanced picture of public perception that directly challenges simplistic narratives.
One of the key discoveries is the 'agency effect.' When an AI system had moderate agency (human sets goal, AI determines means) or high agency (AI sets both goal and means), participants attributed greater causal responsibility to the AI itself. However, under low AI agency—where a human specifies both the goal and the method—blame shifted decisively back to the human, despite their temporal distance from the bad outcome. This suggests perceived autonomy is a major driver of blame.
The research uncovered other critical biases. Participants consistently judged a human as more causal than an AI, even when both performed identical actions in reversed-role scenarios. Perhaps most impactful for the industry, AI developers were judged as highly causal actors in the chain, which reduced blame on end-users but did not reduce blame on the AI system itself. Furthermore, when an AI was decomposed into a large language model (LLM) and an agentic component, the public placed more blame on the agentic part that takes actions.
- Participants blamed AI more when it had moderate or high agency (autonomy in goals/means), but blamed humans more under low-agency setups.
- AI developers were consistently judged as highly causal, reducing blame on human users but not on the AI system itself.
- In decomposed systems, the agentic component (the part that takes actions) was blamed more than the underlying LLM model.
Why It Matters
This research provides the empirical foundation needed to design fair liability frameworks and policies for real-world AI incidents.