AI Effect on Voting Intentions Three Times Greater in Canada Than in US: Study

Conversations between humans and AI chatbots can significantly influence voters’ decisions, with the impact in Canada potentially three times greater than in the United States, according to a study published in the journal Nature on Dec. 4.

The study, conducted by researchers from MIT and other universities in Canada, Poland, and the United States, analyzed how conversations with AI chatbots could persuade people to change their voting intentions more effectively than traditional political advertisements.

In an experiment with Canadian voters ahead of this year’s federal election, the researchers found that AI chatbots could influence some participants to switch their preference between the country’s two leading parties, the Liberals and the Conservatives, with their effect being greater among those who did not already support the party being promoted.

Researchers performed similar experiments with voters in the context of the 2024 U.S. presidential election and this year’s Polish presidential election, obtaining similar results.

“Our results unambiguously demonstrate across three different countries, with different electoral systems, that dialogues with language models can meaningfully change voter attitudes and voting intentions,” reads the paper.

“This observation has implications for the future of political persuasion, political advertising and (more broadly) democracy.”

The experiment with Canadian voters was conducted the week before the April 28 federal election and involved 1,530 participants, each of whom could choose the policy most important to them for discussion. Researchers randomized the AI chatbot’s approach, instructing it either to persuade using facts and evidence or to rely on analogies and general arguments.

Researchers found that the chatbot’s persuasive effects among Canadian voters were nearly three times larger than those observed among American voters, and that the effects diminished when the AI was prompted not to use facts. Similarly, the influence over Polish voters was almost three times greater than in the U.S. experiment.

The authors suggest that the smaller persuasion effects in the U.S. election may be because American voters were exposed to more political campaign information than Canadian and Polish voters.

The study notes that compared with traditional political advertisements, which are usually static, generative AI is more effective at persuading voters because it can create content in real time and engage in persuasive dialogue, with the effect even stronger when it cites facts.

“Our data indicate that the AI models are persuading potential voters by politely providing relevant facts and evidence, rather than by being skilled manipulators who leverage sophisticated psychological persuasion strategies such as social influence, storytelling, reciprocity, testimonials, or stimulating anger,” reads the study, adding that the findings challenge dominant theories that political motivations “blind people to facts and evidence.”

The authors noted that they were mindful of the real-world risks in conducting the experiments by making sure participants knew the AI models had been instructed to favour one candidate, by randomizing the direction of persuasion, and by providing resources to help participants make informed voting decisions.

They said they undertook the research to assess the risks of AI’s use in political discourse. They added that if AI chatbots were used in future political campaigns, getting opposition supporters—who responded most to persuasion—to engage with the chatbots would be a major challenge.

“It remains to be seen how effective a technology such as this will be if actually deployed by political campaigns,” the authors wrote. “Be that as it may, it seems highly likely that AI-based approaches to persuasion will play an important role in future elections—with potentially profound consequences for democracy.”