Proceedings of the Annual Meeting of the Cognitive Science Society

Much of high-level cognition relies on identifying which possibilities are relevant in a situation, and representations of available alternative actions can predict the extent to which people attribute blame. Building on work showing that large language models (LLMs) sample option spaces differently from humans, we examine how moral evaluations change when potential alternative actions are supplied by an LLM rather than generated by a human. In Study 1, considering LLM-generated alternative options led participants to evaluate the agent's chosen action more negatively and to attribute more blame than when participants generated options themselves. In Study 2, participants rated LLM-generated options as having a higher value than human-generated ones and again attributed more blame after considering LLM-generated options. We additionally show that LLM-generated options occupied a narrower region of the semantic space than human-generated options. The implications for AI's influence on human cognition and judgment are discussed.