The Pulse
AI Answers Are Becoming a Midterm Voter’s First Stop
AI-generated answers and search summaries are already shaping how voters find election information, sometimes surfacing misleading pages or incomplete answers. Companies are adding official voting resources, but studies and reporting show t
AI.info Team · Written with AI · fact-checked ·
“We’re campaigning for the AI vote now,”
Eric Wilson, director of the Center for Campaign Innovation, told The Atlantic.
That campaign is unfolding through answers voters see before they reach a campaign website or election office. In a report published October 22, The Atlantic documented how AI-generated search results and chatbots are already mediating information about candidates, donations and voting. The examples show a new point of vulnerability: a confident answer can lend authority to a bad link or a dubious source.
A donation link sent money to the wrong campaign
Several people searching for ways to donate to Texas Democratic Senate candidate James Talarico encountered a Google AI answer that linked to a fundraising page bearing his name and photograph. The page, however, directed donations to a political-action committee that sent the money to Colorado Senator John Hickenlooper’s reelection campaign. Hickenlooper’s campaign said an outside group created the fundraiser without his knowledge; it said at least three people gave a combined $1,011.
The page’s fine print disclosed where the money would go, but Google’s answer highlighted it without catching that distinction. After a Democratic digital-advertising strategist found the page, it was removed from ActBlue, and Hickenlooper’s campaign forwarded the misdirected money to Talarico. The incident was not simply a chatbot inventing a fact: it was a misleading page amplified by an AI answer that failed to surface the qualification that mattered.
AI summaries already reach more voters than chatbots
The Atlantic cited a September survey by the American Association of Political Consultants Foundation in which 4 percent of respondents said they had consulted a chatbot for election information in the previous month. Nearly 40 percent said they had received such information through Google’s AI search summaries. The gap suggests that the voter-information layer is not limited to people deliberately opening ChatGPT or another assistant; it also appears inside familiar search results.
A States United Democracy Center study offers a measure of the accuracy problem. Testing nearly 1,000 AI-generated responses about six states in fall 2025 and winter 2026, researchers found that ChatGPT and Google AI, taken together, directed users to their state’s official election website less than half the time. ChatGPT did so in 39.4 percent of responses; Google AI did so in 55.6 percent. The researchers also found incomplete candidate lists in many responses, even as factual error rates in their tests fell between the two testing periods.
Thin local coverage leaves room for planted sources
Researchers at the cybersecurity start-up VisionHeight told The Atlantic they had found 37 suspicious websites, registered on the same September night, that appeared tailored to questions about state and local candidates. One of the sites recommended a Michigan congressional candidate in response to a question about public safety. Google’s Gemini-powered AI Mode later cited two sites in that network when answering a question about the candidate’s position; the answer was not necessarily wrong, but the sites’ news-like names concealed their nature.
The sites’ operators were not identified. The Atlantic reported that Google’s AI answer changed after the publication asked the company about the network. That sequence does not establish who created the sites or why, but it shows how candidate information gaps can give unfamiliar pages a route into generated answers.
Companies are adding official sources, with limits
OpenAI said in an October 20 update that ChatGPT now integrates voting information from Democracy Works, including registration, deadlines and polling locations, with links to official sources. It also said the Associated Press integration will provide live vote counts after polls close. Anthropic’s midterm safeguards announcement describes state-specific voting answers drawn from Democracy Works’ collection of state and local election information, alongside links directing users to TurboVote.
Those steps address practical voting questions, but they do not settle how systems handle campaign claims, local races or fast-moving rumors. The Brennan Center’s tests of six widely used AI products found that chatbots often pushed back against common election falsehoods, while also making mistakes; the researchers found that models could readily generate misleading election images and video. A model that gives a useful answer about a polling place may still misread political material it retrieves elsewhere.
The test on November 3 will be more specific than whether chatbots sound neutral: when voters ask where to vote, who is running or whether a claim is true, do they get a complete answer grounded in accountable sources? For now, even the basic safeguard of linking to a state election website is inconsistent—and a search result has already helped send real campaign donations to the wrong destination.