AI may act alone, but humans still take the blame

Artificial intelligence may be making decisions on its own, but when things go wrong, humans are still being held responsible.

A new study titled "Human Attribution of Causality to AI Across Agency, Misuse, and Misalignment" examines how people assign blame in complex AI-driven scenarios. The research shows that responsibility rarely shifts fully to machines, even when they act independently.

The researchers conducted a series of experiments to explore how ordinary people interpret responsibility in different AI-related scenarios, focusing on how varying levels of AI autonomy influence perceptions of causality and blame. The results reveal a nuanced and sometimes counterintuitive picture of human judgment in the age of intelligent machines.

Human responsibility persists despite rising AI autonomy

The study found that people continue to assign significant responsibility to humans even when AI systems exhibit high levels of autonomy. This challenges the widely discussed idea of a "responsibility gap," where increasing machine autonomy could lead to reduced human accountability.

In scenarios where AI systems acted independently to achieve a goal using harmful methods, participants often recognized the AI as a key causal agent. However, they still showed a strong tendency to attribute blame and responsibility to human actors, particularly when those humans initiated or influenced the system's actions.

This pattern was especially evident in cases involving low levels of AI agency, where humans retained control over both goals and execution. In these situations, participants overwhelmingly assigned greater causality, blame, and foreseeability to the human agent, even when the AI played a direct role in producing the harmful outcome.

The persistence of human responsibility suggests that people rely heavily on intuitive notions of intention, control, and moral accountability when evaluating complex technological systems. Even as AI systems take on more sophisticated roles, human involvement remains a central factor in public judgments of responsibility.

The findings indicate that legal and regulatory frameworks cannot assume that increased AI autonomy will automatically shift responsibility away from humans. Instead, public expectations may continue to hold human actors accountable, regardless of the technical capabilities of AI systems.

AI agency influences but doesn't determine blame

The study also finds that AI autonomy significantly shapes perceptions of causality, but does not fully determine how blame is assigned. When AI systems were described as having moderate or high levels of agency, participants were more likely to view them as causal contributors to harmful outcomes.

In scenarios where AI systems independently selected harmful means to achieve a goal, participants often rated the AI as more causal, more blameworthy, and more foreseeable than the human user.

However, this increased attribution to AI did not eliminate human responsibility. Even when both the human and AI performed similar actions, participants frequently judged the human as more responsible overall. This suggests that perceptions of responsibility are influenced not only by the actions taken, but also by deeper assumptions about human intention and moral agency.

The research highlights the importance of agency and autonomy as key factors in causal attribution. Systems that are perceived as acting independently and making decisions without direct human control are more likely to be seen as causal agents.

The study also shows that these perceptions are not straightforward. In some cases, participants attributed greater responsibility to the AI, particularly when it played a direct and visible role in producing harm. In other cases, responsibility shifted back to the human, even when the AI exhibited similar behavior.

This variability reflects the complexity of human reasoning in situations involving multiple interacting agents. People do not rely on a single rule to assign responsibility; instead, they consider a range of factors, including intention, control, proximity to the outcome, and perceived autonomy.

Developers emerge as critical but overlooked actors

The study identifies developers as a key but often underappreciated source of responsibility in AI-related harms. When developers were included in the causal chain, participants frequently assigned them a significant degree of causality, even though they were more distant from the immediate outcome.

The presence of a developer in the scenario had a notable effect on how responsibility was distributed. Participants tended to reduce the level of responsibility assigned to the human user when a developer was involved, suggesting that accountability shifts depending on how the causal chain is structured.

However, the inclusion of developers did not significantly reduce the responsibility attributed to the AI system itself. This indicates that people may view developers and AI systems as contributing to harm in different ways, rather than as interchangeable sources of responsibility.

The study also explores how breaking down AI systems into components affects perception. When the AI was decomposed into a large language model and an agentic tool, participants assigned greater causal responsibility to the agentic component. This suggests that people distinguish between different types of AI functions, attributing more responsibility to parts of the system that appear to act with greater autonomy.

These findings have important implications for the design of accountability frameworks. As AI systems become more complex, involving multiple layers of technology and human input, determining responsibility will require a more nuanced understanding of how different actors contribute to outcomes.

Rethinking legal and policy frameworks for AI accountability

Current approaches to AI accountability may not fully align with how people intuitively assign responsibility. Legal systems often rely on clear causal relationships to determine liability, but AI systems introduce layers of complexity that make such determinations more difficult.

In real-world scenarios, harmful outcomes often result from a combination of human decisions, system behavior, and broader institutional factors. This creates what the study describes as a "dense causal web," where multiple actors contribute to the final outcome.

The research suggests that understanding public perceptions of causality is critical for developing effective legal frameworks. Since many legal judgments are influenced by common-sense reasoning, insights into how people assign responsibility can help inform policy decisions.

Policymakers should avoid assuming that AI systems will absorb responsibility as they become more autonomous. Instead, responsibility may continue to be distributed across humans, developers, and AI systems in complex and context-dependent ways.

The findings also highlight the importance of transparency in AI systems. Clear information about how systems operate, who controls them, and how decisions are made can help stakeholders better understand responsibility and accountability.