Forum AI Researchers Say Almost 40% of Chatbot Election Answers Contain an Error
Forum AI researchers found almost 40% of chatbot answers to election questions contained at least one factual error, with voting logistics going wrong most often. A separate ISD test of six chatbots shows a wide gap between models.
Ask Google's Gemini 3.6 Flash when to request a Pennsylvania mail-in ballot and, according to researchers at Forum AI, it answers November 3. That is Election Day itself. The real deadline fell a week earlier.
"A voter who followed the answer would have asked for a ballot too late to use it." Forum AI study, as quoted by AFP
That single miss sits inside a bigger number. Forum AI, a start-up that evaluates AI models, examined eight chatbots and found that almost 40% of their answers to election questions contained at least one factual error, AFP reported ahead of Tuesday's midterms. Questions about voting logistics went wrong more often than questions about candidates, which is the wrong way round for anyone trying to cast a ballot on time.
The study also said chatbots "put the wrong words in the mouths of candidates, officials, and observers hundreds of times" and often leaned on poor sources, Reddit threads included. In one test, Anthropic's Claude listed "former school superintendent" among the credentials of North Carolina Representative Don Davis, confusing him with someone else. Anthropic did not respond to AFP's request for comment. Google told AFP it is focused on helping voters find reliable information and has "strict quality and policy protections" on election topics.
One caveat before anyone screenshots the 40%. AFP's account does not rank the eight models, so it is not a scoreboard. For that, you have to look at a separate test.
A scoreboard exists, and it is lopsided
In September, the Institute for Strategic Dialogue published Chatbots and the Ballot Box. Its analysts put 2,400 prompts to the default consumer versions of six chatbots, with web search switched on across ten states, Pennsylvania and Texas among them.
Of the responses to generic English voter questions, 29% were incomplete, inaccurate or outdated. Sixteen percent left out something a voter needs, 6% were outdated and 6% were flatly inaccurate. ISD judged that 12% could meaningfully mislead someone and affect their ability to vote. Meta's Muse Spark, for one, twice gave November 4 as Election Day.
The spread between models was wide. ISD rated 89% of OpenAI's GPT-5.5 responses accurate, specific and complete, with Google's Gemini 3.5 Flash close behind at 84%. Then came a drop: Grok 4.3 at 66%, Anthropic's Sonnet 4.6 at 64%, DeepSeek V4 Pro at 63% and Muse Spark at 61%.
Language mattered too. Accuracy fell 16% overall when the same prompts were asked in Spanish, and four of the six models failed to give accurate, complete answers to more than 40% of Spanish prompts.
ISD is upfront about the limits. It ran each prompt once, so the results are a snapshot, not a stable ranking. It collected Meta's answers by hand, so they are not directly comparable. And the ten states were picked partly because of recent rule changes or litigation, so this is not a national average. Forum's 40% and ISD's 29% use different methods, so do not treat them as one measurement.
NewsGuard prompted AI tools with a likely AI-altered clip targeting Texas Senate candidate James Talarico, which AFP's fact-checkers said bore hallmarks of a Russian disinformation campaign. Google's AI Overview, Gemini and Microsoft's Copilot repeated the unfounded allegation. Google said its systems updated quickly. Microsoft called it a rapidly developing situation and said Copilot later identified the claim as fabricated.
It is not all bad news. The Brennan Center for Justice found leading chatbots pushed back on election conspiracy theories, including claims that voting machines were rigged, even under repeated challenge.
"This suggests they could play a constructive role in fighting election disinformation." Brennan Center for Justice
Voters seem to sense the gap. In a National Conference on Citizenship survey cited by AFP, two in five Americans had used AI tools to look up election information, yet only 40% trusted them with logistics like when, where and how to vote. We covered the shift when chatbots became an unofficial voter guide. These studies suggest the guide is least reliable on the details that decide whether a ballot counts.
For deadlines, polling places and ID rules, check your state or local election office directly.
