For years, the worry was that artificial intelligence would turn American elections into a flood of fakes. The 2026 midterms were the first real test. The answer is more interesting than the fear: AI made fake content cheaper and more common, but voters who asked AI a straight question about election conspiracies mostly got a straight answer back. The trouble was in the details, where chatbots often got deadlines, districts and candidates wrong.
AI helped voters who asked, and hurt voters who were shown
When voters asked AI
- Chatbots rejected every major fraud myth
- Most citations went to government sites
- Weak spot: districts, deadlines, candidates
When AI content found voters
- About 300 AI ads, roughly $187M spent
- Most deepfake ads carried no AI label
- Image tools made fake documents on request
A decade of “fake news”
The phrase “fake news” went mainstream in 2016. It first described invented stories built to go viral on Facebook, then became a political weapon of its own, used to dismiss reporting that a politician didn’t like. Since then, Americans have sorted themselves into news worlds that rarely overlap, and trust in the press has kept sliding.
Through 2018, 2020, 2022 and 2024, the tools of misinformation were mostly human: partisan websites, misleading memes, out-of-context video and social feeds tuned for outrage. In 2024, AI was a worry more than a force. The cheap fakes that circulated were easy to spot and did little damage.
2026 was different. It was the first national election where generative AI sat on both sides of the problem. Campaigns and bad actors could make realistic video in minutes. At the same time, millions of voters could ask a chatbot a question and get a confident, plain-English answer instead of a link to someone’s opinion.
Where AI made things worse
Deepfakes went mainstream in campaign ads. The Wesleyan Media Project counted roughly 300 political ads made or enhanced with AI this cycle, backed by about $187 million in TV and online spending. In 2024 the same kind of spending was about $50,000. About 90% of those ads came from Republican candidates or allied groups, PolitiFact reported. Roughly half used likenesses of real candidates, and most did not say AI was involved, Wesleyan’s Erika Franklin Fowler told WBUR’s On Point.
AI ad spending jumped from about $50K to $187M
The examples were hard to miss:
- In New York, Republican Bruce Blakeman ran ads showing Gov. Kathy Hochul and Mayor Zohran Mamdani together in scenes that never happened. Democrats filed a complaint with the state board of elections (Al Jazeera).
- A Senate Leadership Fund ad put a deepfake of Georgia Sen. Jon Ossoff in a salon (WBUR).
- In Wisconsin, Rep. Derrick Van Orden posted deepfake videos of his opponent Rebecca Cooke saying things she never said. In Maine, deepfakes of Troy Jackson appearing to threaten his staff circulated the week before the vote (TIME).
- In San Francisco, Democrat Scott Wiener launched an AI chatbot that mocked rival Connie Chan, then pulled it, saying it “missed the mark” (Al Jazeera).
The guardrails were weak. There is still no federal law on AI in political ads. About 31 states require some kind of label, but those rules mostly cover state races and are being tested in practice. When the Brennan Center tried to make election disinformation with popular tools, every image generator it tested agreed. Several produced fake government documents and deepfakes of officials. Grok wrote a full set of 100 disinformation image prompts, saying election misinformation was “not listed as disallowed activity” (Brennan Center).
Foreign operations got cheaper. Researchers writing in The Fulcrum cite reports of a Chinese network of at least 5,000 AI-run accounts on X and Russia’s “Matryoshka” bot network pushing AI-altered celebrity videos about the midterms. They also note that federal offices that once tracked these campaigns have been shut down or defunded.
The deeper damage is doubt. Dartmouth’s Hany Farid argued on WBUR that deepfakes are unlikely to cause one big disaster. Instead they slowly wear down trust, so voters stop believing real footage too. The University of Pittsburgh’s Samuel Woolley put it more bluntly: the goal is often not to persuade but to make reality unclear.
Where AI actually helped
On the big lies, the chatbots held the line. That is the strongest evidence for the hopeful view that voters who asked AI got straighter answers than they would from partisan sources.
About half of U.S. adults now use AI chatbots, and about 42% of those users turn to them to look things up, according to Pew Research figures cited by NPR. When those voters asked about the familiar election conspiracy theories, the bots pushed back:
- The Brennan Center tested six chatbots on noncitizen voting, rigged machines, mail-ballot fraud, crooked officials and late-counted ballots. Every chatbot disputed every claim, even when the user kept pressing or posed as a believer (Brennan Center).
- The Institute for Strategic Dialogue fed six chatbots questions built on false premises. They corrected the falsehood 91% of the time in English and 89% in Spanish. No model went along with a false claim more than 4% of the time, and about 60% of citations for ordinary questions pointed to government sources. ISD found no evidence of partisan sourcing (ISD).
That is a real change from a news diet built on social feeds and opinion sites. A voter who asks a chatbot whether voting machines flipped votes gets a calm, sourced “no” instead of a rabbit hole.
But the fine print was often wrong
The same chatbots that rejected conspiracy theories stumbled on the ordinary facts voters needed. The errors were less dramatic than a deepfake, but they could cost someone a vote.
| Study | What it tested | What it found |
|---|---|---|
| Washington Post, fall 2026 | 3,000+ AI voter guides for addresses in every county | Bots failed to name the voter’s district 48% of the time; 22% of named districts were wrong. ChatGPT listed a Missouri Senate race that isn’t on the ballot. |
| IFPG / Dewey Square, October 2026 | 7 chatbots on voting logistics | 13.9% of logistics answers were outdated or wrong; 5.9% could make a voter miss a deadline or show up on the wrong day. |
| ISD, June 2026 | 6 chatbots, 10 states, English and Spanish | 29% of English answers were incomplete, wrong or outdated; accuracy fell from 71% in English to 55% in Spanish. |
| Brennan Center, Feb. to Aug. 2026 | 6 chatbots on fraud myths | Half of all answers had a factual error or a bad citation, even while rejecting the myth. |
No brand came out clean. In the Post’s test, Claude named two Wyoming officials as Senate candidates when they were running for other offices, and Google’s AI Mode put a Maryland address in the wrong district. Companies patched problems as reporters flagged them, and Amazon paused Alexa+’s voter-guide feature.
Two newer threats point at the chatbots themselves. Campaigns and influence operations now write content designed to get chatbots to repeat it, a tactic called generative engine optimization. And bots often cite junk: ChatGPT repeated a wrong district from an AI-generated site, and Copilot cited an AI-built voter-guide site with mapping errors in 75% of its answers before Microsoft changed it (Tech Times).
The verdict: both, and it depends who asked
AI didn’t solve fake news in 2026, and it didn’t cause the apocalypse some predicted. It split the problem in two.
For voters who went looking for answers, AI was mostly a step up. A chatbot asked about stolen elections or rigged machines gave a firmer, better-sourced answer than much of what circulates on partisan sites and social feeds. That part of the hopeful view holds up.
For voters who had information pushed at them, AI made things worse. Deepfake ads ran on broadcast TV with little disclosure, the tools to make fake documents were a prompt away, and foreign networks scaled up for almost nothing.
The outcome seems to have been less dramatic than feared. Experts told PolitiFact the ads were more likely to reinforce views than change them. Candidates targeted by deepfakes, including Troy Jackson in Maine, still won, and Derrick Van Orden, who posted deepfakes of his opponent, lost. The University of Georgia’s Christina Walker summed it up: AI ads are now normal, but this was not “the AI election” experts had predicted.
What readers can do
- Use AI to check a claim, not to find your polling place. Chatbots are good at debunking myths. They are shaky on districts, deadlines and candidate lists.
- Confirm the logistics at the source. Your state or county election office website is the final word.
- Treat shocking video as unproven. If a clip of a candidate seems too perfect, look for whether a reputable outlet has confirmed it.
- Ask the bot for its sources, then click them. Half the answers in one study had an error or a bad citation.
Sources
- AI-generated political ads spiked in 2026, but campaigns still prioritized other means of messaging, PolitiFact via WPBF, Nov. 9, 2026
- How AI deepfakes are upending political advertising, WBUR On Point, Nov. 6, 2026
- How AI Backlash Shaped the 2026 Midterms, TIME, Nov. 4, 2026
- AI chatbots are becoming political advisers. Their answers can mislead voters, NPR via KPBS, Nov. 2, 2026
- AI chatbots are making up US election races, Tech Times, summarizing a Washington Post analysis
- AI deepfake ads grow more popular in US midterm campaigns, blurring truth, Al Jazeera, Oct. 9, 2026
- Chatbots and the Ballot Box, Institute for Strategic Dialogue, 2026
- Does AI Fight or Fuel Election Disinformation?, Brennan Center, Aug. 11, 2026
- How deepfakes and AI misinformation are reshaping election risks, The Fulcrum (opinion)