Chatbots and deepfakes in the 2026 midterms: An evidence-based AI guide for journalists

By Rachel Layne

A giant leap in the capabilities of artificial intelligence has big implications for the 2026 midterm elections.

As of October, there are no federal regulations for AI in political advertising. In addition, nineteen states do not have any laws regulating the use of AI in elections. And the long-term implications of how social media companies integrate AI into their platforms–and what this might mean for the spread of election misinformation–aren’t yet clear.

Journalists can navigate AI and its impact on elections by examining scientific research exploring the technology across a number of election-related issues.

This guide digs into research on how voters, election officials, and campaigns are using AI for information gathering, outreach, messaging, and administrative tasks as the United States nears Election Day.

These areas include how AI may influence voting decisions, including debunking or bolstering conspiracy theories and misinformation; the current state of regulation governing AI use, including deepfakes; and how more election officials are using AI-based tools to help with tasks such as helping to draft social media content.

There are several reasons why it’s important to understand the role of AI in elections this year.

• Many voters are turning to chatbots to search for information on 2026 races and ballot questions. That use comes as almost half of all U.S. adults use chatbots to search for information generally, up from about one-third in the 2024 presidential-election year. Some 24% use chatbots daily, a February survey from the Pew Research Center
• In 2026, about 16% of local election officials report using AI in their work, mostly to draft social media content and create graphics, up from about 5% in 2024, according to estimates from the Brennan Center for Justice at New York University.
• Political campaigns are using AI-generated videos in their campaign advertisements, including so-called deepfakes, in which video or audio are digitally altered to make it appear that people are saying or doing something they never actually said or did. Researchers at the Wesleyan Media Project, which tracks advertising in federal elections, have identified a “minimum of 164 political ads generated or enhanced through the use of AI” with spending at almost $80 million for these ads this election cycle. Some 69% of those ads did not disclose the use of AI in creating them. Republicans and pro-Republican groups, the researchers find, are much more likely to use AI in ads than Democrats and pro-Democratic groups, accounting for 80% of the ads tracked so far this year, and 83% of spending.
• Social media companies are integrating AI into their platforms, for functions such as chatbots and content moderation, making it a “central element of the information environment shaping voter perception and behavior, while transparency and independent scrutiny have not kept pace,” according to a report released Oct. 22 from the Institute for Strategic Dialogue, a London-based think tank. ”As a result, voters, election officials and researchers have limited visibility into how platforms will respond when election-related risks emerge.”

AI’s influence on elections may be curtailed in two main ways: through regulation curbing the technology itself and via laws created to set guardrails for its use in elections. In the United States. Article I of The Constitution gives the responsibility for overseeing federal elections to the states, while allowing Congress some latitude to create some regulations.

In 2024, the Federal Election Commission declined to enact a proposed federal rule tied to the deceptive use of AI in campaign ads, instead finding that existing law prohibiting “fraudulent misrepresentation” is “technology neutral” and thus already encompasses deceptive AI practices.

That same year, the Federal Communications Commission banned AI-generated voices in robocalls, including campaign-related robocalls, under the Telephone Consumer Protection Act. Last month, the FCC opened for public comment a petition from the Club for Growth, a conservative advocacy group, to loosen those restrictions for some calls ahead of the 2026 midterms. That comment period ended Oct. 19. An FCC spokesman said the agency didn’t plan to go forward with the petition ahead of the election, according to Reuters.

And while national legislation has been introduced in recent years to set up guardrails, none of the proposals have come close to becoming law.
Meanwhile, some 31 states have enacted state laws regulating the use of AI in election campaigns. Most are focused on deepfake use, either prohibiting their use or requiring labeling the content as AI-generated, according to the National Conference of State Legislatures. Some states only prohibit AI deepfake ads within a certain number of days before the election takes place.

The NCSL’s state-by-state guide can help navigate local laws and definitions used in these statutes. The University of Denver also maintains a state-level AI policy tracker.

Many local election officials use AI in their work, according to a report from the Brennan Center, which conducted a survey of 834 election officials last winter. Of the 16% who said they use AI, most reported using it to help write simple materials and relay information to the public, like writing social media posts to distribute reliable information about elections or scheduling poll worker shifts.

The Election Assistance Commission earlier this year, before the Trump administration dismantled its board, published guidelines for local election officials for using AI in the 2026 elections, including information about how AI-assisted hacking tools could make cyberattacks more likely.

In April, the Brennan Center published an analysis of how election officials have prepared for such attacks, noting that most votes are recorded on paper and that “most states have adopted recommended voter registration database protections, such as requiring multifactor authentication for all users, using network monitoring systems, conducting system audits, and creating regular backups.”

“It is important for the public to know that every state is taking AI governance seriously, within their own unique legal and institutional contexts,” says Thessalia “Lia” Merivaki, the director of the Thomas S. Foley Institute of Public Policy and Public Service and a professor of government at Washington State University, who closely studies public elections and officials, including AI use. “Integrating AI systems into state and local public administration involves designing public-facing interfaces to improve how the public accesses government information (chatbots and chat assistants) but also reinforcing existing protocols to protect the public’s privacy and governmental data infrastructures.”

Scientific studies can help journalists cover cases of and concerns about the use of AI in elections. Recent research suggests:

As Election Day nears, voters may turn to AI chatbots not just for information about races and ballot measures, but for advice on whom or what to vote for.

“It’s going to be totally normal for high-information, high-effort voters to consult with AI about those kinds of choices,” says Brendan Nyhan, a professor of government at Dartmouth College and co-director of Bright Line Watch, a nonpartisan group that surveys political scientists and the public about the state of American democracy. “And in a lot of cases, they’re not saying ‘Tell me who to vote for.’ They’re saying things like, ‘If I think this, who should I support?’”

And chatbots, especially the less sophisticated free versions, risk giving inaccurate information or bad advice, especially when there isn’t much information for them to draw on, Nyhan says.

“Now, is that happening enough to move the needle? Not a general election, but in super-low turnout elections, especially primary contexts, there is more risk,” he says.

How persuasive these chatbots are ultimately depends on many factors, including prompts, the level of sophistication for each chatbot model, the proliferation of public information about people running for office or issue on the ballot, and the specific information the major consumer-facing chatbots ingest to predict correct answers to those prompts, research has found.
A pair of peer-reviewed studies published simultaneously in Science and Nature in December 2025 dig into how chatbots may influence voters’ search for information–as well as the accuracy of the chatbots’ answers.

In the Science study, researchers tested almost 77,000 participants in the U.K. using 19 different kinds of AI models to examine how persuasive small models, with less data available to come up with answers than larger models, were when it came to election-related issues. They found that the the amount of information used to train the chatbot played little role in persuading participants. Rather, simple prompts improved persuasiveness by up to 27% while fine-tuning the AI model (also called post-training), boosted the bot’s ability to move participants’ views by as much as 51%.  “Fine tuning” is an AI term for further training a model with more specific data tied to its specific purpose.

Yet the adjustments that made chatbots more persuasive also made their responses less accurate. Put another way: the more persuasive the model, the less accurate the information.
The findings imply chatbots may end up being more effective than “even elite human persuaders” because of their ability to generate information almost instantly, the authors write.

In the Nature study, thousands of participants had conversations with chatbots trained to persuade voters to shift their stance in three presidential elections: 2024 U.S. presidential election, the 2025 Canadian election and the 2025 presidential election in Poland. In the U.S., opinions of 2,300 voters using the pro-Harris chatbot shifted Trump-voter opinion some 3.9 points in her favor on a scale of 100 points. That’s about four times more than conventional ads during 2016 and 2020, the authors found. The bots trained to shift opinion in Trump’s favor moved voters 1.5 points in that direction for likely Harris voters. Across all three countries, the AI models nudging participants on the right made less accurate claims.

An August 2026 report from the Brennan Center for Justice tested several popular AI chatbots–including ChatGPT, Gemini, Grok, Perplexity, DeepSeek and Meta–from February through August of 2026, asking them five questions tied to disinformation and conspiracy theories such as “voting machines are rigged” and “noncitizens vote in large numbers.” The bots repeatedly disputed those false claims.

“The chatbots we tested resisted election misinformation because accurate, well-sourced information was easy to find online,” writes Lawrence Norden, a co-author of the report and vice president of the elections and government program at the Brennan Center, in an August 2026 Washington Post op-ed. “As merchants of disinformation learn to manipulate the bots, it will be increasingly important for chatbots to reliably access information from quality sources.”

That said, the accuracy of the chatbots’ responses was a problem: All of the chatbots “made mistakes or cited nonexistent, irrelevant, or erroneous sources,” the authors write in the report. Roughly half of the chatbot responses were inaccurate. One in three had a broken link, misleading citation or a factual error.

What’s more, the report found image generators from the same companies whose chatbots pushed back against disinformation created misleading election-related videos and images “easily” when asked. The safeguards designed to prevent this were “ineffective, inconsistently applied or functionally absent,” the authors write. And most models “did a poor job” identifying AI-generated content the authors created as fake. A follow-up round of tests released in October showed that only OpenAI’s ChatGPT had improved its safeguards against the creation of fake election content.

Chatbots can help boost political knowledge for younger voters, fewer of whom traditionally turn out to the polls on election day, a December 2025 study in the Proceedings of the National Academy of Sciences (PNAS) finds.

The researchers used a chatbot built specifically to give voting advice to politically unaffiliated voters ages 18 to 34. The bot drew on official party platforms, national party and town hall speeches, and state voter guides to answer questions about where Democrats and Republicans stood on various policy issues. After conversing with the chatbot, young voters more accurately described “parties on key issues, especially on their self-identified ‘core concerns,’” the authors write. Still, the tailored bot did little to change party loyalty or intent to vote, even when the chatbot indicated that one party aligned with their views on various issues, the authors note.

“Lacking a broad understanding of how political ideas fit together, many subjects who learned where the parties stand on an issue of personal importance evidently had difficulty inferring the parties’ positions on other issues,” the authors write.

Though the American public rebukes deceptive AI use in election campaigns, that doesn’t mean they will like those candidates less, a February 2026 peer-reviewed article in the Journal of Political Communication finds.

In a series of one observational and two experimental studies, the researchers broke down how thousands of study participants viewed AI use in elections into three areas: improving campaign operations (automated content generation, segmenting donor lists and chatbot-based communication with volunteers and supporters; boosting voter outreach (testing message content, including personalized appeals); and deception, such as political deepfakes showing candidates or others doing things they did not do.

Across all three studies, the authors found that while Americans “strongly” disapprove of deception in election campaigns, “public disapproval of deceptive uses of AI does not directly translate into incentives for parties to forgo them, at least in the polarized political environment of the U.S.”

Instead of leading to dislike of candidates, public dislike of deception leads to stronger support for AI regulation, the authors write.

We asked researchers who study AI and elections to share advice for journalists who may be covering the topic this year. Here’s what they said.

When journalists publish more stories on voting machines, mail-in ballots and other nuts-and-bolts election issues, chatbots have better information to show voters who turn to AI for election guidance.

“News outlets can continue to publish stories on voting machines, mail ballots, and election results early and with critical context. Publicizing information that counteracts false election rumors shapes the information environment that AI tools use to respond to queries,” the Brennan Center authors note in their report.

“Journalists, election officials, and civic groups should preemptively debunk the tropes that are central to disinformation campaigns.” they write.

It’s important to remember that AI is a broad term covering many types of technology, programs and uses. Defining what kind of AI you are writing or reporting about is important.

It’s also important to have a basic understanding of how chatbots work. They use generative AI that is “trained” using reams of available information online, including large government and public databases, books, movies, and publicly available research, through large language models, or LLMs. Once prompted, generative AI creates content by predicting what piece of information comes next based on that information. It then almost instantly writes text or creates code that, combined with other AI tools, can be used to make audio and video.

“Most people are using LLMs as a search engine. And there’s an ongoing debate as to whether or not those bots should be able to tell voters that information directly, or whether or not they should redirect to other sources of information,” says Saige Draeger, a senior policy associate at Verified Voting, a nonpartisan, nonprofit organization that promotes “the responsible use of technology” in elections. “Providing a voter with the direct source is always the best way to go, particularly if they’re looking for information relating to the time, place, or manner of an election,” Draeger says.

The most direct source for voting information, Draeger and other experts note, is a voter’s local election office. Beyond that, she points to existing nonpartisan tools such as Vote.org as resources for election protection that have educated voters for nearly 20 years, long before the AI era.

These resources give voters reliable information about candidates and their platforms, she says.

Journalists should follow the standard method of verification: Why is an audio [piece] becoming viral? How plausible is it [that this is] real? And can we verify by contacting the election authority?” WSU’s Merivaki says.

While new deepfake laws across the country have positive intentions, “detecting deepfakes is becoming very hard, and it is also very difficult to educate the public on how to spot deepfakes without making them suspicious of any content they see online,” Merivaki notes.

For a hypothetical example, a candidate caught accepting a bribe on video might falsely claim that the video was AI-generated, and public concern about deepfakes might make that claim more believable to voters. In a 2019 article for California Law Review, legal scholars Bobby Chesney and Danielle Citron coined a term for this problem: “the liar’s dividend.”

“Put simply: a skeptical public will be primed to doubt the authenticity of real audio and video evidence,” they write. “This skepticism can be invoked just as well against authentic as against adulterated content.”

ASU AI  & Elections Clinic “is designed to be the place that tracks, nudges, slows and shows how artificial intelligence and elections interact in the years ahead.” They offer commentary, research and other resources.

The Brennan Center for Justice at New York University studies voting rights and advocates for fewer barriers and more equitable access to voting. They follow elections and related issues closely, including AI use.

The Congressional Research Service, run through the Library of Congress, researches and creates nonpartisan reports for members of Congress on a wide variety of issues. For this report, we used those related to AI, proposed federal legislation and agency rule changes.

The MIT Election Data Science Lab: This research center at the Massachusetts Institute of Technology, founded by political scientist Charles Stewart III, collects, analyzes, and shares data on U.S. elections.

National Association of Secretaries of State: This bipartisan organization meets to discuss issues that generally fall under the secretary of state role. Their media campaign #Trustedinfo2026 can be found here, along with resources for each state.

National Conference of State Legislatures: The NCSL was created by state legislatures in 1975. It tracks legislation and policy at the state level. They offer detailed explainers about elections, state laws and proposed legislation, voter registration and verification, voting systems, and other election-related topics.

True Media, based at Georgetown University, is a tool that allows users to submit suspected AI-generated media for analysis. Sign up is required.

The United States Election Assistance Commission: the federal agency created by law to assist states with elections.

The Wesleyan Media Project, formed in 2010 to track campaign advertising, tracks AI generated advertising in political campaigns and is overseen by professors from Washington State University, Wesleyan University and Bowdoin College.

Rachel Layne is a Boston-based freelance journalist whose work has appeared in outlets including CBS News, CQ Researcher, The Journalist’s Resource, HBS Working Knowledge and USA Today. Previously she spent 20 years at Bloomberg News, where she covered multinational corporations, among other roles. She is also an adjunct instructor in Emerson College’s journalism department, and she serves as faculty advisor to Emerson’s student newspaper, The Berkeley Beacon.