New study: AI causal attribution models fail to match human blame

A new arXiv paper from researchers Nripsuta Ani Saxena, Stelios Triantafyllou, and Goran Radanović tackles a growing problem: when AI systems make high-stakes decisions, who is responsible for failures? The team compared formal models of responsibility attribution—grounded in actual causality theory—against how humans actually assign blame. Using a modified version of the card game Goofspiel, they created multi-agent sequential decision-making scenarios where multiple AI agents (and human players) contributed to outcomes.

After conducting a large-scale survey to collect human judgments, the researchers evaluated several responsibility attribution methods. Their key finding: no single formal method consistently aligned with human responses. Instead, agent-specific biases and the amount of information available to agents during decision-making emerged as major factors shaping responsibility judgments. This suggests current causal attribution frameworks are not yet robust enough for real-world accountability applications like auditing autonomous systems or determining liability in AI-caused accidents. The paper is available on arXiv (2608.04318).