AI chatbots may have the power to influence voters’ opinions
Enrique Shore / Alamy
Does the persuasive power of AI chatbots spell the beginning of the end for democracy? In one of the largest surveys to date exploring how these tools can influence voter attitudes, AI chatbots were more persuasive than traditional political campaign tools including advertisements and pamphlets, and as persuasive as seasoned political campaigners. But at least some researchers identify reasons for optimism in the way in which the AI tools shifted opinions.
We have already seen that AI chatbots like ChatGPT can be highly convincing, persuading conspiracy theorists that their beliefs are incorrect and winning more support for a viewpoint when pitted against human debaters. This persuasive power has naturally led to fears that AI could place its digital thumb on the scale in consequential elections, or that bad actors could marshal these chatbots to steer users towards their preferred political candidates.
The bad news is that these fears may not be totally baseless. In a study of thousands of voters taking part in recent US, Canadian and Polish presidential elections, David Rand at the Massachusetts Institute of Technology and his colleagues found that AI chatbots were surprisingly effective at convincing people to vote for a particular candidate or change their support for a particular issue.
“Even for attitudes about presidential candidates, which are thought to be these very hard-to-move and solidified attitudes, the conversations with these models can have much bigger effects than you would expect based on previous work,” says Rand.
For the US election tests, Rand and his team asked 2400 voters to indicate either what their most important policy issue was or to name the personal characteristic of a potential president that was most important to them. Each voter was then asked to rate on a 100-point scale their preference for the two leading candidates – Donald Trump and Kamala Harris – and provide written answers to questions that aimed to understand why they held these preferences.
These answers were then fed into an AI chatbot, such as ChatGPT, and the bot was tasked either with convincing the voter to increase support and voting likelihood for the candidate they favoured or with convincing them to support the unfavoured candidate. The chatbot did this through a dialogue totalling about 6 minutes, consisting of three questions and responses.
In assessments after the AI interactions, and in follow-ups a month later, Rand and his team found that people changed their answers by an average of about 2.9 points for political candidates.
The researchers also explored the AI’s ability to change opinions on specific policies. They found that the AI could change voters’ opinions on the legalisation of psychedelics – making the voter either more or less likely to favour the move – by about 10 points. Video advertisements only shifted the dial about 4.5 points, and text ads moved it only 2.25 points.
The size of these effects is surprising, says Sacha Altay at the University of Zurich, Switzerland. “Compared to classic political campaigns and political persuasion, the effects that they report in the papers are much bigger and more similar to what you find when you have experts talking with people one on one,” says Altay.
A more encouraging finding from the work, however, is that these persuasions were largely because of the deployment of factual arguments, rather than from personalisation, which focuses on targeting information at a user based on personal information about them that the user might not be aware has been made available to political operatives.
In a separate study of nearly 77,000 people in the UK, testing 19 large language models on 707 different political issues, Rand and his colleagues found that the AIs were most persuasive when they used factual claims and less so when they tried to personalise their arguments for a particular person.
“It’s essentially just making compelling arguments that causes people to shift their opinions,” says Rand.
“It’s good news for democracy,” says Altay. “It means people can be swayed by facts and opinions more than personalisation or manipulation techniques.”
It will be important to replicate these results with more research, says Claes de Vreese at the University of Amsterdam in the Netherlands. But even if they are replicated, the artificial environments of these studies, where people were asked to interact at length with chatbots, might be very different to how people encounter AI in the real world, he says.
“If you put people in an experimental setting and ask them to, in a highly concentrated fashion, have an interaction about politics, then that differs slightly from how most of us interact with politics, either with friends or peers or not at all,” he says.
That being said, we are increasingly seeing evidence that people are using AI chatbots for political voting advice, according to de Vreese. A recent survey of more than a thousand Dutch voters for the 2025 national elections found that around 1 in 10 people would consult an AI for advice on political candidates, parties or election issues. “That’s not insignificant, especially when elections are becoming closer,” says de Vreese.
Even if people don’t have extended interactions with chatbots, however, the insertion of AI into the political process is unavoidable, says de Vreese, from politicians asking the tools for policy advice to AI writing political ads. “We have to come to terms with the fact that, as both researchers and as societies, generative AI is now an integral part of our election process,” he says.
Topics:
