Short conversations with AI-driven chatbots can meaningfully alter political attitudes, according to two new studies published in Nature and Science. The authors report that under certain conditions, a single interaction with a bot can shift voter preferences by as much as 25 percentage points.
The research was conducted by Professors David Rand and Gordon Pennycook of Cornell University. They found that chatbots are often more persuasive than traditional political advertising, largely because they generate personalized, fact-based arguments in real time. At the same time, the most persuasive models were also more prone to producing inaccurate information — a combination that raises concerns about their potential impact on democratic processes.
In the study published in Nature, the researchers analyzed the behavior of more than 2,300 Americans, 1,530 Canadians, and 2,118 Poles. Participants who interacted with chatbots favoring Kamala Harris shifted 3.9 points toward her on a 100-point scale. Strikingly, even Harris supporters who were exposed to Trump-leaning chatbots shifted 1.51 points toward Donald Trump.
In Canada and Poland, the effect was even stronger — reaching around 10 percentage points — suggesting that under real campaign conditions, AI could meaningfully sway election outcomes. The authors emphasize that even brief conversations can have measurable influence.
The second study, published in Science, involved nearly 77,000 people in the United Kingdom. Participants interacted with 19 different AI models on more than 700 political topics. The most optimized model was able to shift opposition voters by up to 25 points.
Professor Rand admitted that the magnitude of the effect surprised him. He added that such a powerful influence, especially in the context of presidential politics, raises serious questions about the future of campaigning. He noted that the persuasive impact of AI is likely to grow as the technology continues to advance.
The studies also showed that AI’s persuasive strength depends primarily on the number of facts presented. Chatbots instructed to “densify” their arguments with factual content increased their effectiveness by 27%. Models trained with specialized techniques achieved gains of up to 51%.
When researchers restricted the models’ access to factual data, their persuasive power dropped sharply. This indicates that the key factor is not psychological manipulation but the intensity and precision of argumentation. The findings underscore that regulating the use of AI in politics is becoming an urgent issue.
Source: https://ceo.com.pl/naukowcy-ostrzegaja-chatboty-moga-realnie-zmieniac-wyniki-wyborow-37537
