AI chatbots can persuade voters

Source: Chiara Vargiu, Alessandro Nai — Nature News & Views (on Lin et al. and Hackenburg et al.) · 2025-12-04

Conversations with AI can sway people’s political views. Concerningly, a chatbot’s facts are not always accurate, especially when it supports right-wing positions. See p.394

Introduction. The world is getting used to ‘talking’ to machines. Technology that just months ago seemed improbable or marginal has erupted quickly into the everyday lives of millions, perhaps billions, of people. Generative conversational artificial-intelligence systems, such as OpenAI’s ChatGPT, are being used to optimize tasks, plan holidays and seek advice on matters ranging from the trivial to the existential — a quiet exchange of words that is shaping how decisions are made1. Against this backdrop, the urgent question is: can the same conversational skills that make AI into helpful assistants also turn them into powerful political actors? In a pair of studies2,3 in Nature and Science, researchers show that dialogues with large language models (LLMs) can shift people’s attitudes towards political candidates and policy issues. The researchers also identify which features of conversational AI systems make them persuasive, and what risks they might pose for democracy. Political scientists have long debated whether election campaigns have the power to change minds. The answer to this question depends on the specifics of what is being investigated, but a realistic estimate is that political advertising shifts vote choices by less than one percentage point4. Although such modest shifts can matter a great deal in close races, findings from Lin et al.2 (page 394) and Hackenburg et al.3 suggest that AI-driven persuasion operates on a different scale. Lin and colleagues report the results of a controlled experiment in which more than 2,300 participants in the United States chatted with an AI model that tried to convince them to support either Kamala Harris or Donald Trump in the 2024 US presidential election (Fig. 1a). Participants exchanged three rounds of messages with the model, which was instructed to remain respectful and fact-based while tailoring its arguments to each participant’s stated political preferences. Similar experiments were run in Massachusetts on a ballot initiative to legalize psychedelic drugs, and later in Canada and Poland around their 2025 national elections.

Method. Models that were fine-tuned specifically for persuasion were up to 50% more effective than their untuned counterparts. Hackenburg and colleagues analysed a maximally persuasive configuration that ‘pulls all the levers at once’, including personalizing messages. They estimated that it could shift opinions by 16 points on average and 26 points among participants who initially disagreed with a given issue. However, there was a catch: the same settings also reduced factual accuracy. Notably, information-dense models were the most convincing, but also the most error-prone. Persuasion and accuracy, it seems, trade off against each other. It is important to note that these findings come from controlled online experiments. It is unclear how such persuasive effects would play out in real political environments in which exposure to persuasive AI agents is (often) voluntary and conscious. Such environments also contain a myriad of contrasting messages competing for attention, and users can ultimately decide to avoid or ignore specific information sources. Nevertheless, these studies provide compelling systematic evidence that conversational

Discussion. AI systems hold the power, or at least the potential, to shape political attitudes across diverse contexts. The ability to respond to users conversationally could make such systems uniquely powerful political actors, much more influential than conventional campaign media. At the same time, the results highlight that inaccuracy seems to be a crucial downside. The AI models in these experiments were instructed to remain factual, and still produced distortions. In real-world use, where incentives might favour engagement rather than truth, the line between effective persuasion and misinformation could blur further. The broader implication is clear: persuasion is no longer a uniquely ‘human’ business. As AI systems become embedded in chat platforms, virtual assistants and social-media feeds, political dialogue between voters and machines could be an increasing occurrence. Crucially, if conversation with AI can move opinions more than most campaign tools can, there is a fundamental question that will need substantive answers from a legal, ethical and philosophical standpoint: who controls this capacity? Regulations should address not only what AI systems say, but also how they are trained and optimized. Transparency about post-training objectives, disclosure when users are conversing with AI and limits on political outreach will probably become essential components of

Conclusion. Finally, and perhaps optimistically, these experiments offer insights that extend beyond AI, pointing at the state of political discourse in general. People were most persuadable when the AI reasoned politely and spoke in evidence-based terms — qualities that are often seen as lacking in human political debate6. And although the chatbots were instructed to be civil and seemed to comply with the injunction — something much less likely to occur in the real world — a hopeful lesson is that civility and facts still matter. At least when they are delivered by a machine.