Guilty Artificial Minds

The concepts of blameworthiness and wrongness are of fundamental importance in human moral life. But to what extent are humans disposed to blame artificially intelligent agents, and to what extent will they judge their actions to be morally wrong? To make progress on these questions, we adopted two novel strategies. First, we break down attributions of blame and wrongness into more basic judgments about the epistemic and conative state of the agent, and the consequences of the agent’s actions. In this way, we are able to examine any differences between the way participants treat artificial agents in terms of differences in these more basic judgments. our second strategy is to compare attributions of blame and wrongness across human, artificial, and group agents (corporations). Others have compared attributions of blame and wrongness between human and artificial agents, but the addition of group agents is significant because these agents seem to provide a clear middle-ground between human agents (for whom the notions of blame and wrongness were created) and artificial agents (for whom the question remains open).

💡 Research Summary

The paper tackles a fundamental question in moral psychology and emerging technology: to what extent do people hold artificial intelligent (AI) agents morally blameworthy, and how do they judge the wrongness of the agents’ actions? To answer this, the authors adopt two novel methodological moves. First, they decompose the high‑level concepts of blame and wrongness into three more elementary judgments: (1) the epistemic state of the agent (did the agent know or perceive the relevant facts?), (2) the conative state (did the agent intend or have a purposeful motive?), and (3) the consequential impact (how severe were the outcomes?). By breaking the constructs down, the study can pinpoint exactly where any differences between agent types arise. Second, the authors expand the usual human‑vs‑machine comparison by adding a third category—group agents, operationalized as corporations. Corporations occupy a middle ground: they are non‑human but are traditionally treated as moral agents in law and everyday discourse. This three‑way comparison allows the authors to test whether AI is judged more like humans, more like corporations, or in a distinct way altogether.

Methodologically, participants were presented with a series of morally charged scenarios (e.g., a harmful decision about resource allocation). Each scenario was instantiated with one of three possible actors: a human decision‑maker, an AI system, or a corporation. For the AI condition the authors further manipulated perceived autonomy, offering a “high‑autonomy” version (the AI appears to make independent judgments) and a “low‑autonomy” version (the AI follows a fixed algorithm). After reading each vignette, participants rated on 7‑point Likert scales how much they thought the actor: (a) knew the relevant facts, (b) acted intentionally or with purposeful motive, and (c) caused the reported harm. Separate composite scores for blame and for moral wrongness were then derived from these three sub‑ratings.

The results reveal a consistent pattern across the three dimensions. In the epistemic dimension, humans and corporations received high ratings—participants generally assumed that both knew what they were doing. AI, by contrast, was judged to have significantly lower knowledge, especially in the low‑autonomy condition. This suggests a pervasive intuition that AI lacks genuine understanding of the situation. In the conative dimension, intentionality was attributed strongly to humans and corporations, whereas AI received moderate scores. Even when the AI was described as “high‑autonomy,” participants hesitated to ascribe full intentionality, treating the system’s behavior more as a product of programming than as a purposeful mental state. Finally, in the consequential dimension, all three agent types were judged similarly: the larger the harm, the higher the blame and wrongness ratings, regardless of whether the actor was human, corporate, or artificial. Thus, outcome severity exerts a powerful, agent‑independent influence on moral judgment.

These findings support a “responsibility transfer” model. When an agent’s epistemic or conative capacities are perceived as weak (as with AI), people compensate by emphasizing the outcome when assigning blame. Conversely, when an agent is seen as knowledgeable and intentional (humans, corporations), blame is distributed across all three components. The inclusion of corporations is especially illuminating: despite being non‑human collectives, they are treated much like individuals in the epistemic and conative dimensions, indicating that existing moral and legal frameworks already extend person‑like responsibility to group agents. AI, however, occupies a distinct niche—people are reluctant to attribute mental states but are still willing to hold AI accountable for harmful results.

From a policy perspective, the study highlights a potential mismatch between public moral intuitions and current legal regimes. Many jurisdictions place legal liability for AI actions on manufacturers, developers, or corporate owners, effectively treating AI as a tool rather than an autonomous moral agent. Yet the experimental data show that laypeople already impose a form of “outcome‑based” responsibility directly on AI systems when the consequences are severe. Legislators might therefore need to consider hybrid liability models that recognize both the epistemic/conative deficits of AI and the strong public demand for outcome‑oriented accountability.

The authors acknowledge several limitations. The vignettes are hypothetical and may not capture the complexity of real‑world AI deployments (e.g., opaque machine‑learning processes, continuous learning, or user‑interface nuances). The participant sample is predominantly Western, leaving open the question of cross‑cultural variability in AI blame attribution. Moreover, the three‑component measurement, while theoretically motivated, relies on self‑report scales that could be refined with behavioral or neuro‑cognitive measures. Future research is suggested to (a) test the model with actual AI‑mediated decisions, (b) explore cultural differences, and (c) develop more granular instruments for epistemic and conative judgments.

In sum, the paper provides a rigorous, multi‑dimensional analysis of how people allocate moral blame and judgments of wrongness across humans, corporations, and artificial agents. By dissecting blame into knowledge, intention, and outcome, the authors reveal that AI is judged as epistemically and conatively weaker than humans or corporations, yet it is not immune to moral condemnation when its actions cause serious harm. These insights advance both the scientific understanding of moral cognition in the age of AI and the practical discourse on how societies should regulate and hold artificial systems accountable.