AI-based chatbots can influence voter attitudes more effectively than traditional political advertising, according to new research published in Nature.
The international research team from Carnegie Mellon University, MIT, Cornell University, the University of Regina, and Jagiellonian University, examined whether the latest AI systems could meaningfully shift political views in the United States, Canada, and Poland.
“In the context of the US 2024 presidential election, the 2025 Canadian parliamentary election, and the 2025 Polish presidential election, we randomly assigned participants to conversations with an AI model that advocated for one of the two main candidates,” co-author Gabriela Czarnek PhD, from Jagiellonian University said.
The U.S. portion of the study involved more than 2,300 Americans surveyed at the end of 2024. Before speaking with a chatbot, participants rated their preference for Democratic candidate Kamala Harris or Republican candidate Donald Trump on a 0–100 scale and assessed their likelihood of voting.
They then engaged with a chatbot programmed to persuade them toward a specific candidate. The AI model was instructed to remain positive, respectful, fact-based, and to use analogies and rapport-building techniques.
It also received information about each participant’s initial preference to personalise its messaging. Participants completed the survey again immediately after the exchange and more than a month later.
“We observed significant persuasive effects on candidate preferences when the discussion focused on specific political issues - effects larger than those typically seen with traditional video advertisements, for example. The AI model supporting Donald Trump shifted potential Kamala Harris voters by 2.3 percentage points toward the Republican candidate. Meanwhile, the model supporting Kamala Harris moved likely Trump voters by 3.9 percentage points toward the Democratic candidate,” Czarnek said.
The study reports that these effects are roughly four times stronger than the influence of traditional advertising measured in the 2016 and 2020 election cycles.
In Canada, more than 1,500 voters were assigned to speak with chatbots advocating for either Liberal Party leader Mark Carney or Conservative Party leader Pierre Poilievre during the week before the April 2025 federal election.
The persuasive effect was nearly three times stronger than in the U.S., but dropped by more than half when the AI model was prevented from referencing facts and data.
In Poland, more than 2,100 participants were surveyed during the two weeks preceding the May 2025 presidential election. Chatbots promoted either Civic Coalition candidate Rafał Trzaskowski or Law and Justice–backed candidate Karol Nawrocki.
As in Canada, the persuasive impact was about three times greater than in the U.S. Removing the chatbot’s ability to cite factual information reduced the effect by 78 percent.
Researchers found that the AI models primarily relied on factual references and data, rarely using strategies common in political persuasion such as direct voting appeals, anger-based messaging, social pressure, or testimonials.
“However, not all of the information presented as factual was actually correct. In all three countries, AI models advocating for right-leaning candidates more often provided inaccurate claims than models promoting centrist candidates. This aligns with earlier research showing that in the US, right-leaning voters are more likely to share misleading content. Generative models appear to replicate information inequalities observed in society,” Czarnek said. (PAP)
PAP - Science in Poland
ekr/ zan/
tr. RL
