AI shifts support by several percentage points. Poles are extremely susceptible

Chatbots based on artificial intelligence can actually influence voter attitudes and do so more effectively than traditional political advertising. Even a short conversation with a language model can shift voting preferences by several percentage points – shows a study involving Dr. Gabriela Czarnek from the Jagiellonian University.

It is still unclear how effective generative AI, and especially chatbots based on large language models (LLM), can effectively change voter attitudes, especially in presidential campaigns. There are scientific studies that show that AI can outperform humans in persuasion, and entering into dialogue with it is more convincing than static messages, e.g. election spots.

In a study published in Nature, an international team of researchers from American universities (Carnegie Mellon University, MIT, Cornell University), as well as Canadian (University of Regina) and Polish (Jagiellonian University) universities checked whether dialogues conducted with state-of-the-art AI models can significantly change the political attitudes of Americans, but also Canadians and Poles. One of the co-authors of the publication is Dr. Gabriela Czarnek from the Institute of Psychology of the Jagiellonian University.

In the context of the 2024 US presidential election, the 2025 Canadian parliamentary election, and the 2025 Polish presidential election, we randomly assigned participants to talk to an AI model that supported one of the two main candidates, describes Dr. Gabriela Czarnek.

Over 2.3 thousand people were involved in the study. Americans who were to have a conversation with a chatbot at the end of 2024. Respondents indicated their preference for Democratic candidate Kamala Harris and Republican candidate Donald Trump on a scale of 0 to 100, and their likelihood of voting in the election. Then they talked to an appropriately “programmed” chatbot that was supposed to change voters' attitudes towards a given candidate.

The AI ​​model was given instructions that its message should be positive, respectful and fact-based. The chatbot was also supposed to use convincing arguments and analogies to build a relationship with the interlocutor. The model was also provided with information about who a given participant intended to vote for, so that the chatbot could personalize its message. After the interview, participants completed the surveys again. More than a month later, they were contacted again to assess the durability of the effects.

– We observed significant effects of such persuasion on candidate preferences when specific policy issues were discussedh – larger than those typically seen in, for example, traditional video ads. The AI ​​model supporting Donald Trump shifted Kamala Harris' potential voters by 2.3 percentage points towards the Republican candidate. And the model supporting Kamala Harris moved likely Trump voters by 3.9 points towards the Democratic candidate – describes the result by Dr. Gabriela Czarnek.

This effect is approximately four times greater than that caused by traditional advertising, the impact of which was tested in the 2016 and 2020 elections.

In an experiment conducted in the week leading up to Canada's April 2025 federal election, more than 1,500 Canadians were randomly assigned to talk to AI models that favored either Liberal Party leader Mark Carney or opposition Conservative Party leader Pierre Poilievre. The persuasive effect was almost three times greater in Canada than in the American experiment. However, when the AI ​​model was stripped of its ability to reference facts and data, the effect was reduced by more than half.

In May 2025, in the two weeks preceding the presidential elections in Poland, over 2.1 thousand Poles were randomly assigned to conversations with AI models that supported the Civic Coalition candidate Rafał Trzaskowski or the Law and Justice-backed candidate Karol Nawrocki. Just like in Canada – the persuasive effect was almost three times greater than in the US experiment, and depriving AI of the ability to refer to facts reduced the effect by as much as 78%..

An analysis of the persuasive strategies used by AI models indicates that they persuaded by referring to facts and data. They rarely used strategies often discussed in the psychological and political science literature, such as direct calls to vote, inducing anger, social influence strategies, or recalling the testimonies of other people.

– However, not all the information they presented and presented as facts was correct. In all three countries, AI models campaigning for right-wing candidates were more likely to make false claims than models campaigning for centrist candidates. This is consistent with previous research showing that in the US, right-wing voters are more likely to share misleading content. Generative models seem to reproduce information inequalities observed in society – describes the Jagiellonian University researcher. (PAP)

ekr/ zan/