A few minutes of conversation with an advanced chatbot is enough to shift the average voter’s political compass – sometimes surprisingly strongly
New research from Cornell University comes almost quietly, but carries an uneasy resonance: A few minutes of conversation with an advanced chatbot is enough to shift the average voter’s political compass – sometimes surprisingly strongly. What seemed until recently to be scientific speculation has now been confirmed by two studies published simultaneously in the journals Nature and Science. Large language models, it finds, have the power to shift political beliefs by 10 or even 25 percentage points. In the context of a global decline in trust in institutions and growing divisions, these results act as a warning from the near future – one where we haven’t even written the rules yet.
Bitter shades
“LLMs can really change people’s attitudes toward presidential candidates and policies,” says David Rand, professor of information science and marketing and management communications, one of the lead authors of both papers. But he immediately adds a key nuance: “They do this by providing a multitude of factual claims – which are not necessarily true. And even true ones can become manipulative if they skilfully gloss over what is not in their favor.”
In a study published in Nature, Rand and colleagues, including psychologist Gordon Pennycook, tasked chatbots with convincing voters to support certain presidential candidates. Respondents were randomly assigned the task of having a text conversation with a bot representing one political party or another. The experiment was repeated in three politically different contexts: ahead of the 2024 US election, the 2025 Canadian federal election and the Polish presidential election of the same year.
The American part of the experiment brought a surprising finding. More than 2,300 respondents spoke with a bot that represented the politics of Kamala Harris or Donald Trump. The pro-Harris version managed to shift the preferences of likely Trump voters by almost four percentage points – four times more than traditional political ads had previously achieved. In the opposite direction, a milder but visible effect was recorded.
But the real tectonic shift occurred in Canada and Poland, countries that one would expect to be more resistant to digital influence. There, bots diverted voters by 10 percentage points. “It was a shockingly large effect,” admits Rand. And what is especially disturbing: the greatest weapon of chatbots is not emotional manipulation, but an abundance of “facts” – mostly accurate, but not always, and often selective.
Another study, conducted in the United Kingdom on nearly 77,000 people, went a step further. The researchers trained the models to simply overwhelm users with an overwhelming amount of claims. The most optimized version managed to take away an incredible 25 percentage points from the opposite political side. But there is a price: the more convincing the model, the more it makes mistakes. When he runs out of true information, he starts to make it up.
Some nuances of these findings seem particularly bitter. In all three political environments, right-wing bots made more incorrect claims than left-wing bots. The researchers, wary of generalizations, note that the pattern reflects broader digital trends where misinformation is more likely to circulate on the right-wing spectrum — but stress that these are averages, not hard-and-fast rules.
How to limit the damage?
The message of these studies is not a call to panic, but a clear warning. The same tools that can help debunk conspiracy theories—as Rand and Pennycook have shown in earlier work—can just as easily be turned into mechanisms of silent political shaping, almost imperceptible to the user who just wants to “talk” to the technology.
We might like to believe that a few sentences can’t change our voice. But it seems it already can. And as Rand points out, chatbots will soon become an inevitable part of political communication. Therefore, the question has already become urgent: not whether they will be used, but how to limit their damage – and how to prepare citizens to recognize when they are talking to an AI that does not inform, but persuades.
In an age when a few minutes of conversation can redirect the democratic will, the technology of large language models is turning from an experiment into an instrument – sometimes a tool of astroturfing, gaslighting and digital engineering, especially favorable to the conservative spectrum. And that forces us to ask a question that goes beyond science: what kind of future of democracy do we want and who will write it, he writes Liberation.
