Does AI Personalization Lead to Belief Polarization?

Abstract

People increasingly use AI chatbots as a source of information. Many AI assistants use memory of prior conversations to personalize future responses, potentially reinforcing users' existing beliefs. We experimentally study the impact of AI memory on political beliefs in a nationally-representative sample of 1,000 U.S. participants. After discussing Donald Trump's approval on trade and the economy with an AI chatbot, participants forecast the probability that his approval will increase or decrease. We randomly assign participants to either have no AI assistance or to have the option to receive information from an AI with or without memory of their previous discussion. Memory shifts the AI's forecasts towards participants' prior beliefs, but it does not polarize participants' posterior beliefs. Participants with the option to access AI form more moderate beliefs than participants without AI, but memory has a null effect on beliefs. The different effects of memory on AI and participants can be explained by participants' demand for memory: participants are 12 pp more likely to click on the AI forecast when it has memory, and are 21 pp more likely to prefer that their AI forecasts use memory. Since information from the AI has a moderating effect, the increased take-up of AI with memory offsets its sycophantic effects.