Abstract
People increasingly turn to AI chatbots to find information, including information about politicians and political issues. Given evidence that Large Language Models (LLMs) have high persuasive capacities, this has raised concerns about their potential to have undue influence on political attitudes. However, most existing research on AI persuasion explicitly prompts the LLM to persuade the user, which is not representative of how people typically engage with LLMs. Whether undirected political conversations with LLMs shift political attitudes, and whether personalisation amplifies any such effect, remains unclear. We address these questions across two studies. First, in a quota-matched survey of UK adults (N = 2,137), we find that political use of AI is already widespread and its outputs are generally perceived as being politically neutral. Second, in a pre-registered randomised controlled trial (N = 2,081), participants conversed with an LLM about either a politician or political issue, with the LLM not being prompted to persuade. We find that conversations about politicians increased favourability toward them, while conversations about policies modestly reduced support. Conversation personalisation did not amplify these effects. Our findings show that AI chatbots are already a meaningful source of political information in the UK, and that undirected chatbot conversations can influence political attitudes.