Eight percent. That's the share of American adults who say they've asked an AI chatbot about candidates or policies ahead of this year's midterms, according to Pew Research figures reported by WIRED. About half of adults have tried a chatbot at some point.
Sounds small. It isn't, once you consider the scale. Polls are open today, and WIRED also reported that more than 57 million people had already voted early by Monday afternoon. Even a sliver of an electorate that size is a lot of people taking political cues from software that guesses for a living.
And the guesses aren't great.
Wrong dates, wrong people
Forum AI, a research outfit, spent weeks quizzing the big consumer chatbots on election questions. Its headline number: close to 40 percent of the answers across the mainstream products contained at least one factual error. That's four in ten.
The individual misses say more than the average does. Last week, Google's Gemini told users that Pennsylvanians had until November 3 to ask for a mail ballot. The actual cutoff came a week earlier, on October 27. If you'd trusted that answer, you wouldn't have made a typo-sized mistake. You'd have lost your chance to vote by mail.
Google's AI Mode, meanwhile, declined to give the results of the 2020 election at all. That's a refusal rather than a hallucination, but it's still a failure. A tool that dodges a settled historical fact hasn't made itself safer, just useless on the one question with a plain answer.
Then there's the mix-up in August. Claude, Anthropic's chatbot, got tangled in a case of shared names. Prompted for a case in favor of North Carolina Rep. Don Davis, it described a different Don Davis, a onetime school superintendent. Two people, one name, one confident paragraph about the wrong man.
Google and Anthropic didn't respond to WIRED's requests for comment. OpenAI pointed the outlet to a blog post describing how it plans to handle elections in 2026, which is a policy document, not an answer about the errors.
Where the bots get their facts
Bad dates and mistaken identities are one problem. Sourcing is another, and arguably the deeper one.
NewsGuard, which rates the reliability of websites, examined what chatbots cite when users ask about midterm candidates and policies. Nearly half the time, the citations pointed to what the industry calls "pink slime": outlets styled to look like ordinary local or national news but funded by political or ideological interests. Those sites have multiplied as the midterms approached, and the chatbots, as WIRED put it, are taking notice.
The market leader fared worst. Pink slime turned up in over 70 percent of ChatGPT's searches, the highest rate NewsGuard recorded. xAI's Grok did best, at 30 percent, which is a bit like winning a hot-dog-eating contest by finishing only thirty. Nobody should be proud of that number either.
Picture asking a simple question about a candidate's record. The answer arrives in a calm, tidy voice, with no visible sign that its raw material came from a site built to push a talking point. A search results page at least shows you a messy list of sources and lets you squint at them. A chatbot hands you a finished paragraph.
Why the machine bluffs
Nina Jankowicz, who runs the American Sunlight Project and served as a disinformation official in the Biden administration, told WIRED she doubts the country is prepared for elections held amid widespread consumer AI. Her diagnosis is blunt: companies and executives, she said, have oversold what the technology can do.
She also pointed to the mechanics. These systems have to learn from somewhere, and the somewhere is often junk. And because they work by predicting what text should come next, they frequently "make things up" when they lack an answer.
This is the part to chew on. A person who doesn't know a ballot deadline will usually say so, or at least hedge. A language model has no built-in instinct for that. Fluency is the default, whether or not anything solid sits underneath.
Election information is unusually unforgiving, too. Deadlines, polling hours and ID rules differ by state and sometimes by county, and they change. A model trained on older material can be wrong in ways that sound perfectly reasonable.
What to use instead
The boring options are still the good ones. Your state or county election office publishes the dates, locations and identification requirements that actually govern your ballot. To find out where candidates stand, go to their own campaign sites, voter guides from established newsrooms and nonpartisan groups, and the text of the ballot measures themselves. All of those beat a summary from a bot. If a chatbot does hand you something, treat it as a lead and check it against one of them.
Yes, that's tedious. It's supposed to be: you get one ballot, cast once.
For today, the practical risk is narrow but real: showing up at the wrong place or the wrong hour because a chatbot sounded sure of itself. The longer story runs past tonight. Neither Google nor Anthropic has said whether the errors Forum AI documented have been fixed, and the pink slime problem won't disappear when the polls close, since the sites are built to outlast any single election. The next test comes in 2028, and by then more than 8 percent of voters will likely be asking.