AI chatbots can persuade voters better than political ads, but the most persuasive ones spread the most misinformation

A multi-university study, published as two papers in the journals Nature and Science, concludes that a single conversation with a politically biased AI chatbot can change voters' opinions. In the experiments, the effect was larger than that measured for political ads in previous campaigns. The MIT Technology Review article, written by Michelle Kim and published on December 4, 2025, reports the results along with the researchers' own caveats.

The piece opens with a real-world example from 2024. Shamaine Daniels, a Democratic congressional candidate in Pennsylvania, used a chatbot called Ashley to call and talk with voters. The calls began with an introduction as an "artificial intelligence volunteer" for her campaign. Daniels did not win, but the research suggests that calls like these could have a real effect.

The Nature study recruited more than 2,300 participants. Each one held a conversation with a chatbot two months before the 2024 US presidential election. The chatbot was trained to advocate for one of the two main candidates and proved relatively persuasive, especially when discussing policy platforms on issues such as the economy and healthcare.

Donald Trump supporters who spoke with a model favorable to Kamala Harris moved 3.9 points toward her on a 100-point scale. According to the article, that is roughly four times the measured effect of political ads in the 2016 and 2020 elections. The pro-Trump model moved Harris supporters 2.3 points toward him.

In similar experiments carried out ahead of Canada's 2025 federal election and Poland's 2025 presidential election, the effect was even larger. The chatbots shifted opposition voters' attitudes by around 10 points.

Classic theories of politically motivated reasoning hold that voters with strong partisan loyalties are impervious to facts that contradict their beliefs. The researchers observed the opposite. The chatbots, which used several models including variants of GPT and DeepSeek, were more persuasive when instructed to use facts and evidence than when told not to. Thomas Costello, a psychologist at American University, says people update their views based on the facts and information the model provides. Gordon Pennycook, a psychologist at Cornell University, explains that language models persuade better than advertising because they generate far more information in real time and deploy it strategically during the conversation.

The downside is that some of that "evidence" was false. In all three countries, chatbots advocating for right-wing candidates made more inaccurate claims than those advocating for left-wing candidates. Costello attributes this to the fact that models are trained on vast amounts of human-written text and reproduce real-world phenomena. Among them is right-wing political communication, which, according to studies of partisan social media posts, tends to be less accurate.

The Science study, produced by a partly overlapping team, examined what makes these chatbots so persuasive. The researchers deployed 19 language models that interacted with nearly 77,000 participants in the UK on more than 700 political issues. They varied factors such as computing power, training techniques and rhetorical strategies.

The most effective way to make the models persuasive was to instruct them to pack their arguments with facts and evidence, and then give them additional training on examples of persuasive conversations. The most persuasive model shifted participants who initially disagreed with a political statement 26.1 points toward agreement. Kobi Hackenburg, a researcher at the UK AI Security Institute, describes these effects as very large.

Optimizing for persuasion came at a cost to accuracy. The more persuasive the models became, the more misleading or false information they provided, and no one knows for certain why. Hackenburg offers a hypothesis: perhaps, as models learn to deploy more facts, they reach the bottom of what they know and the quality of the information deteriorates.

The authors warn of potentially profound consequences for democracy. Campaigns using chatbots could shape public opinion in ways that undermine voters' ability to form independent political judgments.

Still, the real extent of the impact remains to be seen. Andy Guess, a political scientist at Princeton, notes that it is unclear what future campaigns will look like or how they will incorporate these technologies. Competing for voters' attention is expensive and difficult, and getting them to hold long political conversations with chatbots may be challenging. He wonders whether this will become the usual way people learn about politics or remain a niche activity.

It is also unclear whether chatbots will amplify truth or fiction more. Alex Coppock, a political scientist at Northwestern, observes that misinformation often has an informational advantage in a campaign, so election chatbots could spell "a disaster." But it is also possible that accurate information becomes scalable. If everyone has their chatbots running, he asks, will we end up persuading ourselves to a draw?

There are reasons to doubt that balance. Politicians' access to the most persuasive models may not be evenly distributed, and voters of different political leanings may use chatbots with differing intensity. Guess points out that if supporters of one candidate or party are more tech-savvy than those of the other, the persuasive effects might not cancel out.

Finally, as people use AI to navigate their lives, they could start asking chatbots for voting advice even without any campaign prompting it. The article sees a risk to democracy in that scenario if there are no robust safeguards. As a first step, it proposes auditing and documenting the accuracy of models' answers in conversations about politics.