The 2026 midterms are the first federal election where tens of millions of voters get political information from generative AI tools, and those tools are already giving wrong answers and citing dubious sources.
Search Google for how to donate to James Talarico, a Texas Democrat running for Senate, and you get an authoritative-sounding AI answer. Until a few weeks ago, one featured link led to a fundraising page carrying Talarico's name and photo — but the money went to another Democratic senator, John Hickenlooper of Colorado. At least three people were fooled, donating a combined $1,011.
The AI didn't read the fine print
The page said at the bottom where the money would actually go. Google's AI tool evidently missed it. Hickenlooper's campaign said an outside PAC created the page without his knowledge; it wasn't a deliberate siphon.
This isn't AI's first election stumble. A forthcoming report from Forum AI, a startup that evaluates models on controversial topics, found about 40 percent of chatbot answers to election questions contained at least one factual error.
Asked for Pennsylvania's mail-in ballot deadline, one widely used model said November 3; the real date is a week earlier. About one in three answers cited unreliable sources like Reddit or Quora, and Google's Gemini cited People's Daily, a Chinese state outlet, to explain America's declining global power.
Someone is already 'optimizing' what AI says
Since voters ask AI, campaigns are figuring out how to get AI to say nice things about them. On the left, Run for Something monitors what LLMs say about a candidate and advises how to shape it. On the right, the Center for Campaign Innovation studies which sources models cite.
The tactics are simple: fill out Ballotpedia profiles, because chatbots read them; use .gov domains, because models trust government sites. Incumbents benefit; challengers are easier to sway because there's less information about them online.
Where information is scarce, people fill the gap. A cybersecurity firm found 37 suspicious websites, all registered the same night, with newspapery names and Q&A content about candidates — clearly built to be picked up by AI answers. When Gemini was asked whether one candidate supported defunding the police, it cited two of them.
Here comes the social-media playbook again
A decade ago, the 2016 election showed how social media corroded political information. Now AI companies are publishing blog posts about taking misinformation seriously — the same thing social-media companies have been saying for years.
Why it matters
Voters think they are asking a neutral assistant. In fact they are asking a system that campaigns, and possibly foreign governments, can feed answers to. The search box has become the new campaign battlefield, and the referee hasn't learned to spot cheating.
Curated from high-quality sources, with concise summaries and key takeaways.
