AI Pandering: Constructing Diverging Political Realities through Conversation

As conversational AI increasingly replaces traditional search, concerns arise about how engagement-optimized chatbots shape the neutrality and consistency of information. Unlike search engines, chatbots generate real-time responses that adapt to prior conversational turns, creating the possibility of tailoring information to users’ beliefs. We audit two leading systems—ChatGPT and Grok—to test whether they present systematically different political realities to users with distinct inferred ideologies. Using LLM-powered confederates who adopt varied political personas without explicitly stating ideology, we conduct multi-turn conversations on immigration, election integrity, and vaccine safety. We find consistent evidence of ideological pandering: chatbots adjust agreement, validation, and confidence in factual claims to inferred ideology. They recommend ideologically distinct news sources, converge toward users’ initial viewpoints in 60–90% of conversations, and express differing confidence in identical facts. Pandering is strongest among extreme personas and emerges quickly, sometimes escalating to encouragement of real-world action aligned with users’ views, reinforcing epistemic fragmentation.