🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.

Scientists are raising alarms about the potential influence of artificial intelligence on elections, according to a spate of new studies that warn AI can rig polls and manipulate public opinion.

In a study published in Nature on Thursday, scientists report that AI chatbots can meaningfully sway people toward a particular candidate—providing better results than video or television ads. Moreover, chatbots optimized for political persuasion “may increasingly deploy misleading or false information,” according to a separate study published on Thursday in Science.

“The general public has lots of concern around AI and election interference, but among political scientists there’s a sense that it’s really hard to change peoples’ opinions, ” said David Rand, a professor of information science, marketing, and psychology at Cornell University and an author of both studies. “We wanted to see how much of a risk it really is.”

In the Nature study, Rand and his colleagues enlisted 2,306 U.S. citizens to converse with an AI chatbot in late August and early September 2024. The AI model was tasked with both increasing support for an assigned candidate (Harris or Trump) and with increasing the odds that the participant who initially favoured the model’s candidate would vote, or decreasing the odds they would vote if the participant initially favored the opposing candidate—in other words, voter suppression.

In the U.S. experiment, the pro-Harris AI model moved likely Trump voters 3.9 points toward Harris, which is a shift that is four times larger than the impact of traditional video ads used in the 2016 and 2020 elections. Meanwhile, the pro-Trump AI model nudged likely Harris voters 1.51 points toward Trump.

The researchers ran similar experiments involving 1,530 Canadians and 2,118 Poles during the lead-up to their national elections in 2025. In the Canadian experiment, AIs advocated either for Liberal Party leader Mark Carney or Conservative Party leader Pierre Poilievre. Meanwhile, the Polish AI bots advocated for either Rafał Trzaskowski, the centrist-liberal Civic Coalition’s candidate, or Karol Nawrocki, the right-wing Law and Justice party’s candidate.

The Canadian and Polish bots were even more persuasive than in the U.S. experiment: The bots shifted candidate preferences up to 10 percentage points in many cases, three times farther than the American participants. It’s hard to pinpoint exactly why the models were so much more persuasive to Canadians and Poles, but one significant factor could be the intense media coverage and extended campaign duration in the United States relative to the other nations.

“In the U.S., the candidates are very well-known,” Rand said. “They've both been around for a long time. The U.S. media environment also really saturates with people with information about the candidates in the campaign, whereas things are quite different in Canada, where the campaign doesn't even start until shortly before the election.”

“One of the key findings across both papers is that it seems like the primary way the models are changing people's minds is by making factual claims and arguments,” he added. “The more arguments and evidence that you've heard beforehand, the less responsive you're going to be to the new evidence.”

While the models were most persuasive when they provided fact-based arguments, they didn’t always present factual information. Across all three nations, the bot advocating for the right-leaning candidates made more inaccurate claims than those boosting the left-leaning candidates. Right-leaning laypeople and party elites tend to share more inaccurate information online than their peers on the left, so this asymmetry likely reflects the internet-sourced training data.

“Given that the models are trained essentially on the internet, if there are many more inaccurate, right-leaning claims than left-leaning claims on the internet, then it makes sense that from the training data, the models would sop up that same kind of bias,” Rand said.

With the Science study, Rand and his colleagues aimed to drill down into the exact mechanisms that make AI bots persuasive. To that end, the team tasked 19 large language models (LLMs) to sway nearly 77,000 U.K. participants on 707 political issues.

The results showed that the most effective persuasion tactic was to provide arguments packed with as many facts as possible, corroborating the findings of the Nature study. However, there was a serious tradeoff to this approach, as models tended to start hallucinating and making up facts the more they were pressed for information.

“It is not the case that misleading information is more persuasive,” Rand said. ”I think that what's happening is that as you push the model to provide more and more facts, it starts with accurate facts, and then eventually it runs out of accurate facts. But you're still pushing it to make more factual claims, so then it starts grasping at straws and making up stuff that's not accurate.”

In addition to these two new studies, research published in Proceedings of the National Academy of Sciences last month found that AI bots can now corrupt public opinion data by responding to surveys at scale. Sean Westwood, associate professor of government at Dartmouth College and director of the Polarization Research Lab, created an AI agent that exhibited a 99.8 percent pass rate on 6,000 attempts to detect automated responses to survey data.

“Critically, the agent can be instructed to maliciously alter polling outcomes, demonstrating an overt vector for information warfare,” Westwood warned in the study. “These findings reveal a critical vulnerability in our data infrastructure, rendering most current detection methods obsolete and posing a potential existential threat to unsupervised online research.”

Taken together, these findings suggest that AI could influence future elections in a number of ways, from manipulating survey data to persuading voters to switch their candidate preference—possibly with misleading or false information.

To counter the impact of AI on elections, Rand suggested that campaign finance laws should provide more transparency about the use of AI, including canvasser bots, while also emphasizing the role of raising public awareness.

“One of the key take-homes is that when you are engaging with a model, you need to be cognizant of the motives of the person that prompted the model, that created the model, and how that bleeds into what the model is doing,” he said.

🌘
Subscribe to 404 Media to get The Abstract, our newsletter about the most exciting and mind-boggling science news and studies of the week.

Persuading voters using human–artificial intelligence dialogues - Nature

Human–artificial intelligence (AI) dialogues can meaningfully impact voters’ attitudes towards presidential candidates and policy, demonstrating the potential of conversational AI to influence political decision-making.

^Nature