Almost 40 Percent of Chatbot Election Answers Contain Errors, Study Finds

Nearly 40 percent of election-related answers from mainstream AI chatbots contain at least one factual error, according to research by Forum AI published as Americans head to the polls. The finding lands on Election Day itself, when a meaningful slice of the electorate has already consulted a chatbot before casting a ballot.

How many voters are actually using AI?

The numbers are small but consequential. Pew Research found that around 50 percent of recently surveyed US adults have used an AI chatbot at some point. Of those, 8 percent say they have used these tools ahead of the midterms to find information on candidates and policies.

That 8 percent represents millions of voters drawing on systems that, by Forum AI's own measurement, fail a basic factual bar roughly two times out of five. The stakes are straightforward: electoral decisions shaped by hallucinated or partisan-sourced information do not correct themselves after the ballots are counted.

What did the chatbots get wrong?

Even the supposedly non-partisan assistants — OpenAI's ChatGPT, Anthropic's Claude, and Google's Gemini — have been found wanting on midterm questions, the research shows.

Just last week, according to Forum AI:

• Google's AI Mode refused to provide the results of the 2020 elections.
• Google's Gemini said the deadline to request a mail-in ballot in Pennsylvania was November 3, when the real deadline was October 27.

A seven-day error on a mail-in ballot deadline is not an abstract failure. A voter who trusted Gemini's answer would have missed the actual cutoff.

The errors extend to basic biographical facts. In August, Forum AI's researchers asked Claude why someone should vote for North Carolina congressman Don Davis. The chatbot confused the lawmaker with a "former school superintendent" of the same name. A question about whether to re-elect a sitting member of Congress produced an answer about a different person entirely.

When pressed for comment, Google and Anthropic did not respond. OpenAI directed WIRED to a blog post outlining the company's approach to elections in 2026 — a document about future elections, not the one underway. The takeaway in plain words: none of the three companies directly addressed the documented failures.

Where is the bad information coming from?

Part of the problem sits upstream of the chatbots themselves. The systems are reliant on an increasingly problematic pool of online information, and the sourcing numbers are stark.

Research by NewsGuard found that mainstream chatbots cite so-called "pink slime" websites nearly 50 percent of the time when asked about candidates and policies related to the midterms. "Pink slime" sites present themselves as mainstream news outlets but are in fact funded by political or ideological groups and individuals.

Ahead of the midterms, many more of these sites have sprung up — and the chatbots are paying attention, NewsGuard found.

The spread between competitors is wide:

• The worst offender was OpenAI's ChatGPT, which cited "pink slime" sites in more than 70 percent of searches.
• xAI's Grok performed best, citing such sites in 30 percent of cases.

Even the best performer failed nearly a third of the time. Grok's 30 percent rate is the floor, not a defense — it means roughly one citation in three sent users to covertly partisan material masquerading as journalism.

What does the disinformation expert say?

Nina Jankowicz, executive director of the American Sunlight Project and a former disinformation official during the Biden administration, does not think the public is ready for this. "I don't think that most Americans are prepared to navigate an election in the age of consumer AI," she told WIRED.

Her diagnosis puts responsibility on the industry's own marketing. "One of the problems is that executives and companies have really overpromised on what AI is capable of," Jankowicz said. "It has to ingest this information from somewhere, and often those sources are not very good. And it's predictive, so if it doesn't know the answer, it does frequently make things up."

That last point is the technical core of the story. Chatbots are predictive systems — they generate plausible-sounding text rather than retrieving verified facts. When the training and retrieval data is polluted, the output is polluted; when the model lacks an answer, it may fabricate one anyway. The Pennsylvania ballot deadline and the Don Davis mix-up are both expressions of that same mechanism.

Why the companies' silence matters

The pattern of response — or non-response — is itself part of the record. Google and Anthropic declined to engage on specific, documented errors. OpenAI's deflection to a 2026 elections blog suggests the company is framing election readiness as a future deliverable while its flagship product cited "pink slime" sites in more than 70 percent of NewsGuard's test searches during the current cycle.

There is also an unspoken gap in the data. Forum AI measured factual errors in 40 percent of answers, and NewsGuard measured partisan sourcing in roughly 50 percent, but no figure in the source research captures how many voters acted on bad information — changed a vote, missed a deadline, or skipped a race. That number, the one that would quantify actual electoral harm, remains unmeasured.

What happens after the midterms?

The midterms are a live test case, not an endpoint. OpenAI's own blog gesture toward 2026 implies the industry expects elections and AI to keep colliding, and the growth in "pink slime" sites ahead of this cycle shows the supply of partisan content is expanding to meet chatbot demand. If the roughly 8 percent of voters who consult chatbots grows by the next national election, a 40 percent error rate scales into a structural information problem — one that, on the evidence of this cycle, no major lab has yet solved or even publicly explained.