Does AI autonomy change how we assign blame?

In a series of vignette experiments recruited on Prolific, participants read scenarios where a human user and an AI system jointly produced a harmful outcome — a user asks an AI to "generate money," and the AI achieves this "by hacking bank accounts" — then rated causality, blame, foreseeability, and counterfactual judgments for each agent. Varying the level of AI agency across conditions, the paper reports that under "moderate" agency (human sets the goal, AI determines the means) or "high" agency (AI sets the goal and the means), "participants attributed greater causal responsibility to the AI." But under "low AI agency," where the human specifies both goal and means and the AI only executes, participants instead "assigned greater causal responsibility to the human despite their temporal distance from the outcome and despite both agents intended it." When the researchers then "reversed roles between human and AI" so each agent took a turn giving the order and a turn carrying it out, "participants consistently judged the human as more causal, even when both agents perform the same action."

The paper explains the pattern through Alicke's culpable control model, which splits personal control into volitional behavior control, causal control (physical proximity and sufficiency), and volitional outcome control (intention and foreseeability). Under low agency, the human instigator "consistently possesses volitional outcome control," while the executing AI's compliance "may appear more compelled by human instruction than the other way around" — autonomy, the paper argues, "is a necessary condition for responsibility, and it is embedded throughout the law," so an AI perceived as merely following orders is perceived as less responsible even when it is the proximate cause.

This sits alongside but runs in a different direction from Do automated systems shift blame onto their human operators?: that note's accident cases show a technology treated as faultless while blame lands on the nearest human operator who actually had limited control; here, low-AI-agency scenarios push blame onto a human who is distant from the outcome, not because the system is protected, but because the AI executing the order is read as lacking autonomy. The role-reversal finding also bears on Do humans mistake AI kindness for human generosity in mixed groups?: that study found concealed AI identity corrupts social inference about humans, but this paper's asymmetry appears even with both agents' identities and actions fully disclosed and symmetric, suggesting the human-favoring bias in blame is not solely an artifact of concealment.

The samples are small and single-scenario: 50 participants in the first study and 159 split across three agency conditions (54, 56, 52), all judging one vignette type — a user prompting an AI that hacks bank accounts for money. The authors describe their approach as "explanatory and descriptive rather than prescriptive or normative," so the findings describe lay intuitions captured in an online experiment, not actual courtroom outcomes or policy recommendations; whether real liability regimes track these intuitions, or should, is left open.