Does decomposing AI systems shift where people assign blame?

Beyond the human-versus-AI comparisons that make up most of the paper's experiments, one study decomposes "the AI" itself into two parts and asks which one participants hold more responsible. The abstract states the result plainly: "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." The discussion frames this as a Study 5 finding: "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."

The mechanism the paper proposes is a reversion to proximity once the autonomy asymmetry between human and AI is no longer in play. Earlier studies in the same series find that attributions track perceived autonomy when one agent is human and the other is AI, but when both compared agents are AI components — "sharing the same nature" — that asymmetry drops out, and the proximity heuristic documented in prior causal-chain research reasserts itself: the component doing the proximate, world-touching action (the agentic, tool-using part) draws more blame than the component generating the underlying reasoning (the LLM). The paper notes this echoes findings for all-human chains, where "the one closer to the outcome tends to receive higher causal ratings" (citing Cheung et al., 2024).

This connects to Does role-play distinguish real harm from simulated harm?, which treats played action with real consequences as genuine action in effect. This folk-attribution finding gives an empirical echo of that claim at the level of lay judgment: when asked to assign blame, people also locate it in the tool-using, world-touching layer rather than in the generative model behind it. It also bears on Does human-in-the-loop design actually enable effective oversight?, since if folk and future liability judgments concentrate on the agentic layer, oversight built around only the base model's outputs would be observing the wrong component. The finding sits alongside its sibling note, Does AI autonomy change how we assign blame?, as a further decomposition within the same experiment series rather than a separate study.

The finding comes from a single study (Study 5) using the same financial-hacking vignette as the rest of the paper, and the excerpt does not define "agentic component" architecturally beyond that vignette's framing — it does not say how real engineering teams, auditors, or regulators would draw the line between a system's "agentic" and "language model" parts in practice. At the strength the evidence allows, the result suggests that if liability frameworks come to track lay intuition, scrutiny of agentic systems may concentrate on the tool-using layer rather than the underlying model, but the paper offers this as a single experimental result, not a design recommendation.