Political Ingroup and Outgroup Cues Condition How Large Language Models Refute Conspiracy Theories

Despite emerging literature on conspiracy theories (CTs) in human-AI interactions, whether large language models (LLMs) refute CTs and how political identities affect responses remains unclear. Drawing on AI system documentation, positive politeness theory, and motivated reasoning research, we hypothesize that LLMs refute most CTs but adapt responses to the political identities of the user and actor in the claim. In a preregistered study comprising 135,845 conversations, 57 CTs, 21 user-actor political identity condition combinations, and seven OpenAI LLMs, we find that 82% of responses refute the CT and 20% exhibit motivated responding, i.e., acknowledge the user's perspective and affirm their concern. While refutation is more likely for CTs implicating Democrats or Republicans, motivated responding typically accompanies refutation and is more likely when claims involve the user’s outgroup---for example, when a Republican user inquires about a CT involving Democrats, and the model refutes the claim. Although responses to CTs implicating Republicans vary more by user identity than CTs implicating Democrats, the former attenuates response differences between outgroup and ingroup claims. These findings raise broader questions about what it means for AI models to be capable and reliable if LLMs consistently refute conspiracy theories but do so in politically asymmetric ways.