Recent research indicates that conversational AI can significantly influence an individual’s political views, even after a short interaction. Findings published in the journals Science and Nature suggest that these shifts can last for over a month, prompting concerns about the technology’s impact on elections.
In a U.S.-based study, researchers employed AI systems, including OpenAI’s GPT-4o, to engage with supporters of Donald Trump prior to the 2024 presidential election. After conversing with a bot designed to support Democrat Kamala Harris, participants shifted nearly four points on a 100-point scale in her favor.
The effect was even more pronounced in simulated polls for upcoming elections in Canada and Poland, where voter support changed by as much as ten points. Across all studies, the AI’s most effective approach combined a courteous tone with evidence-based arguments, irrespective of their accuracy.
David Rand, a professor at Cornell University and the senior author of the study, highlighted the magnitude of the effect. “In our simulations, about one in ten voters in Canada and Poland altered their voting intentions. In the U.S., it was roughly one in twenty-five,” he noted.
Importantly, the influence did not dissipate quickly. Follow-up surveys revealed that half of the opinion change persisted after a month in the U.K. study, and one-third remained in the U.S. Rand described such enduring results as unusual in social science research.
A significant finding was that AI models promoting right-leaning candidates often conveyed more inaccurate information. Researchers believe this likely mirrors broader trends of misinformation present in the online data used to train these systems.
The studies involved thousands of informed participants who were aware they were interacting with an AI. Experts caution that more advanced models and real-world applications could enhance this persuasive capability, potentially influencing electoral outcomes in closely contested races.
