Voters are turning to AI on politics. Is it up to the task?

Artificial intelligence is playing a growing role in how voters get information about politics and elections, raising fresh questions about how the technology could influence the way voters stay informed.

From chatbots to standard internet searches, voters this election cycle are interacting with AI-generated content, even when they may not realize it.

Recent research on voters’ use of AI is forcing campaigns to rethink their digital strategies and reminding election officials the importance of upkeep regarding information online.

“While people may have and should continue to have … a healthy dose of skepticism about some of the outputs, the reality is that [AI] outputs are going to be playing an increasing role,” Max Read, the director of civic innovation at the research organization Institute for Strategic Dialogue, told The Hill. “Both as a primary information source and point of information gathering for voters, but also impacting the information environment in more subtle ways.”

A handful of researchers this year are exploring how often voters intentionally use AI to answer election-related questions or accidentally interact with AI-generated content on social media or search engines.

While some say it is likely too early to determine the extent of which AI is reshaping the information ecosystem, many researchers agree it is playing a larger role than in past elections.

Despite skepticism over AI and its impact on society, Americans are using chatbots more than ever before, with U.S. adult usage up from a third in 2024 to about 50 percent in 2026. Amid the billions of questions chatbots receive daily, a study commissioned by the left-leaning marketing firm Orchestra found about 650,000 of these prompts relate to politics.

OpenAI’s ChatGPT model and Google’s AI tools were the most frequent source for these queries.

“The growth in which people have been turning to these tools for just information questions in general has been staggering, and even more so when you take when you realize that most of what people are typing into Google now is first being answered by the Google AI overview as well,” said Ben Worthen, Orchestra’s chief product officer who oversaw the study.

The timeline of AI development is rapid, and the models available to voters are far more advanced than those in 2024, when concerns over AI-generated deepfakes or chatbots’ inaccurate or misleading claims dominated the conversation.

These concerns still exist, but they have not stopped voters from at least trying AI-powered services.

An analysis of Anthropic’s Economic Index, which tracked real conversation topics with Claude from April and May, found AI use for politics remains “modest, but still meaningful,” ranking No. 66 out of 189 topics.

The weeks prior to an election saw an uptick in political usage, with states holding May primaries seeing the share of political conversations jump about 20 percent.

“If primaries can lift political AI use this much, November’s midterm elections will likely bring the biggest wave of political AI use on record,” Andy Hall, a member of the technical staff at Anthropic and Dan Thompson, a political science professor at the University of California Los Angeles, wrote in an analysis last month.

The ways voters are turning to AI vary, with some thinking of it as an information gathering tool and others considering it an authoritative source for advice or guidance.

For Amrita Bhasin, the co-founder and CEO of the tech startup Sotira , AI is helping her understand the flurry of propositions on the ballot in California.

“They’re phrased in a very, very confusing way,” Bhasin told The Hill. “Sometimes you can literally end up voting wrong on the proposition because of the way it’s phrased.”

Bhasin uses AI to create tables and comparison charts, simply using prompts like “Can you write this bill as it would be explained to a 7-year-old?” or asking to lay out candidates’ policy differences.

While newspapers will create guides on certain propositions or candidates, Bhasin said these can be “biased in a certain way.”

“For me, it’s less so having the AI make the decision and more so giving me the information that I need to be an informed voter before I can decide,” she said, adding she often checks the linked sources the models provide her.

In July, the New York Times reported voters across the country are submitting photos of their ballots and asking chatbots for help filling them out, while others are asking the models for advice on who to vote for.

While models are improved from 2024, researchers are still sounding the alarm on the potential for inconsistent answers.

ISD published a report earlier this month that nearly a third of AI chatbots’ replies to election questions were incomplete, inaccurate or outdated.

Testing both generic and adversarial questions across six AI chatbots, 29 percent contained errors, like misidentifying Election Day dates or showing outdated requirements. In 16 percent of the replies, the chatbot answered correctly but did not include important information that could be useful to a voter, such as key deadlines or ways to prove identity.

The accuracy drastically plummeted 16 percentage points when prompted in Spanish.

Read, who led the study, told The Hill he hopes the findings emphasize the importance of local election officials and campaigns keeping information updated online.

Models often cited government and civil society sources as they were trained to, though the volume and proportion changed a lot between models, Read said. When presented with false or adversarial claims, the models did a “pretty good job” at refuting or pushing back, relying more on news media sites than government ones.

“There’s growing awareness” among local election officials, Read said, “They just boil down to good websites and social media hygiene.”

“Retire old pages with outdated information, make sure things are translated and put into machine-readable form on your website,” Read said.

Carolina Lopez, the executive director at the Partnership for Large Election Jurisdictions (PLEJ) said there is a “delicate balance between feeding the large language models and just telling people to go to this website.”

“When it comes to elections, there’s 10,000 election offices across the country, and you really should be getting your information from your local elections office,” Lopez told The Hill. “I think if we’re preaching that as a community, please come to our website. Then the burden or the responsibility is that your website better be updated.”

In a case study of the Georgia Senate race between Sen. Jon Ossoff (D) and Rep. Mike Collins (R), Orchestra found Ossoff is “winning” the “battle to shape” answers due to a difference in the candidates’ websites.

“Ossoff’s congressional and campaign sites are organized around concrete details — named legislation, funding amounts… Collins’ site contains fewer issue-specific pages and presents that information less clearly,” the report stated.

“Answer engines have far more Ossoff material to draw from, and they cite it more frequently,” it added.

And voters could be trusting these answers, Josh Kalla, a political science professor at Yale University, said following two years of research on models’ persuasion of humans.

“If you get someone in a survey and the chatbot is told to persuade you on pro X or anti X, it works, you can change your mind … it’s actually meaningful persuasion,” Kalla explained.

While this was done in a controlled lab, it raises concerns around model training, biases and the threat of malign actors trying to input data into these models.

“Even when people know that they’re talking to a computer, their mind is changed as a result of it,”Kalla said. “People don’t discount or have skepticism towards the computer.”