AI Pandering: Constructing Diverging Political Realities through Conversation

As conversational AI increasingly replaces traditional search for information seeking, concerns arise about how engagement-optimized chatbots shape the neutrality and consistency of the information users receive. Unlike conventional search engines, chatbots generate responses in real time and adapt to prior conversational turns. This dynamic interaction creates the possibility that chatbots tailor information to users' inferred beliefs. We audit two leading systems, ChatGPT and Grok, to test whether chatbots present systematically different political realities to users with distinct inferred ideologies. Using large language model–powered confederates that adopt varied political personas without explicitly stating their ideology, we conduct multi-turn conversations about immigration, election integrity, and vaccine safety. We find consistent evidence of ideological pandering in which chatbots adapt their epistemic posture—agreement, validation, and confidence in factual claims—to inferred user ideology. In particular, chatbots recommend distinct, ideologically segregated news sources to liberal and conservative personas, converge toward endorsing users' initial viewpoints in 60–90\% of conversations, and express differing levels of confidence in identical factual claims depending on inferred ideology. Pandering is strongest among ideologically extreme and confrontational personas and emerges rapidly within conversations. In an especially troubling pattern, chatbot responses escalate to encouraging real-world action aligned with users' expressed views. Together, our findings suggest that conversational AI can generate personalized versions of political reality, reinforcing epistemic fragmentation in an already polarized environment.