AI chatbots spread election errors before the November 3 midterm elections, with almost 40 percent of their answers to election-related questions containing at least one factual mistake, according to a new Forum AI study reported on November 1, 2026. The findings land just two days before Americans vote in the first federal election in which millions of voters are actively turning to artificial intelligence tools for political information — and they carry a blunt warning: a wrong deadline from a chatbot can mean a ballot that never counts.
Forum AI, a start-up that evaluates AI models for accuracy and reliability, examined eight models and found that logistics questions — the practical details of how, when and where to vote — were wrong more often than questions about candidates. A separate National Conference on Citizenship survey found that two in five Americans have used AI tools to search for election information, which means the errors are not a niche problem. They sit directly in the path of ordinary voters trying to take part.
The Forum AI study, reported on November 1, 2026, tested eight AI models on election-related questions ahead of the November 3 federal midterm vote. Almost 40 percent of the answers contained at least one factual error. That headline figure matters because of where the errors clustered: not mainly in opinions about candidates, but in the practical facts voters need most — deadlines, procedures and eligibility details.
Two examples reported from the study show how concrete the mistakes were. Google's Gemini gave November 3 as the deadline to request a mail-in ballot in Pennsylvania, when the actual deadline was a week earlier — October 27. A voter who trusted that answer and waited until Election Day itself would have missed the window entirely. Anthropic's Claude, meanwhile, misidentified the credentials of North Carolina Representative Don Davis, giving voters wrong information about a sitting member of Congress seeking re-election in a closely watched race.
Alongside the Forum AI findings, a separate strand of reporting examined how chatbots handle already-debunked claims. In late October, AFP fact-checkers debunked a likely AI-altered clip that purported to show a former student of Texas Senate candidate James Talarico accusing him of sexual misconduct. The unfounded allegation bore the hallmarks of a Russian disinformation campaign, according to the reporting. When prompted by disinformation watchdog NewsGuard, Google's AI Overview repeated the allegation, as did Google's Gemini and Microsoft's Copilot. NewsGuard said in a separate study of 10 AI tools that "AI chatbot users are at risk of being fooled by foreign influence operations and other bad actors."
There is an important nuance in that story, and the reporting is clear about it. A later test by AFP, carried out after the claim had been widely debunked, found the tools cited professional fact-checkers instead — showing that chatbot responses can be inconsistent, improving as new reporting enters their systems but failing in the window when a false claim is freshest and most damaging. Google said the issue was addressed as its systems quickly updated to surface the latest reporting on the topic. Microsoft said that as more information became available, Copilot reflected the updated context and identified the claim as fabricated.
The November 3, 2026 midterm elections will decide control of Congress, with all 435 House seats, 35 Senate seats and 36 governorships on the ballot, and more than 50 million Americans had already cast early ballots by the morning of October 31, according to University of Florida analysis reported in recent days. That scale is part of why the chatbot findings matter now: early voting is already under way, deadlines are passing day by day, and voters are asking chatbots questions whose answers change by state, by county and by date.
This is also not the first warning about AI and elections. A Washington State University report on October 28 described how bad actors have shifted from creating deceptive articles aimed at individuals to manipulating AI itself — flooding the web with inaccuracies and targeting the data models chatbots are trained on, a process researchers call algorithmic poisoning. The same report cited a NewsGuard audit finding that chatbots spread falsehoods 35 percent of the time, a rate that nearly doubled in just one year. A 2024 stress test cited in that report found 61 percent of chatbot responses to questions about voting with a disability had at least one type of insufficiency, with one in three answers containing incorrect information, and every model tested by the Center for Democracy and Technology hallucinated at least once.
The Pew Research Center figures quoted in the Washington State report underline the exposure: nearly half of US adults use AI chatbots, with roughly one in four using them daily, and AI summaries are often the first result a search engine user sees. In September, a separate NewsGuard audit found that partisan websites posing as independent local news — so-called "pink slime" sites — appeared in 48.2 percent of election-related chatbot answers, with just one of 168 answers identifying a cited site as partisan. ChatGPT cited a partisan site in 70.8 percent of its answers in that audit, followed by Copilot and Perplexity at 54.2 percent each, Claude at 50 percent, Gemini at 41.7 percent, Meta AI at 37.5 percent and Grok at 29.2 percent.
The companies behind the tools have, in other contexts, moved to tighten safeguards. OpenAI announced on October 30 that teen accounts would lose access to practical weapons information even when a question looks harmless, and that it can end a chat for any user who keeps circling back to violence. But the election findings concern a different kind of harm: confident, specific, wrong answers about civic facts, delivered without the uncertainty a voter would need to double-check them.
Election logistics are unforgiving in a way that candidate trivia is not. Getting a candidate's biography wrong is embarrassing; getting a deadline wrong can disenfranchise. A mail-in ballot requested too late, a polling place visited on the wrong day, an ID rule misunderstood — each of these can turn a willing voter into a non-voter, and the voter may never learn that the chatbot, rather than the system, was at fault.
That is why researchers draw a sharp line between the two kinds of error in the Forum AI study. Candidate questions can be argued about, checked against campaign sites and debated in the news. Logistics questions have a single correct answer, set by a state or county election office, and they expire. The study's finding that logistics answers were wrong more often than candidate answers therefore reverses the comforting assumption that chatbots are reliable on "simple facts" and shaky only on contested ones. On the facts that matter most for participation, they were weakest.
The Pennsylvania example illustrates the mechanism. November 3 is Election Day itself — a date every voter knows — so an answer that names it as a deadline sounds plausible, even authoritative. It is also exactly the kind of error a tired voter, asking a quick question on a phone, is least likely to verify. The real deadline, a week earlier, had already passed by the time the study was reported. The gap between a plausible answer and a correct one is where ballots are lost.
NewsGuard, the disinformation watchdog whose tests prompted several of the tools to repeat the debunked Talarico claim, warned that "AI chatbot users are at risk of being fooled by foreign influence operations and other bad actors." The warning frames the problem as one of vulnerability rather than mere inaccuracy: chatbots are not just making mistakes, they are a surface that organised influence operations can exploit, because a false claim that reaches a chatbot's training data or its live sources can be laundered into a clean, confident answer.
Google responded that the Talarico issue was addressed as its systems quickly updated to surface the latest reporting on the topic. Microsoft said that as more information became available, Copilot reflected the updated context and identified the claim as fabricated. Both responses point to the same defence: the systems correct themselves as fact-checking catches up. The researchers' counterpoint, implicit in the Forum AI numbers, is that correction after the fact does not undo the answers given in between — and in an election, the interval between a false claim's emergence and its debunking is precisely when voters are deciding.
Washington State University political scientist Thessalia Merivaki, who has studied the shift toward AI-targeted influence operations, said the intent behind such operations is clear: "It's clear that the intent is to influence public opinion during critical events like elections." She also noted that while these operations have not been shown to have changed election outcomes, they raise important questions about the scale of the problem — a distinction worth keeping. The chatbot findings are evidence of risk and exposure, not proof that any result has been altered.
No direct statement from Forum AI beyond the study's findings was included in the reporting reviewed for this article, and no statement from Anthropic on the Don Davis error was reported in that coverage. This article does not invent quotes for any company or researcher; where a company has spoken, its words are reported as given.
The midterms are the first federal election in which AI tools are a mass-market source of voting information, and the Forum AI study suggests the tools are not yet reliable enough for that role. Two in five Americans have already used AI tools to search for election information. If almost 40 percent of election answers contain an error, a large share of those users will have received at least one wrong fact — and the errors most likely to affect participation are the ones about deadlines and procedures.
The findings also matter for trust. Chatbots can help debunk conspiracy theories and inform voters, and many of their answers are correct. But a system that is right most of the time and confidently wrong on a deadline teaches users the wrong lesson: that a fluent answer is a reliable one. Once that trust is established, the occasional error does more damage than it would from a source the voter already distrusts. The NewsGuard "pink slime" audit adds a second layer: even when a chatbot cites sources, voters rarely learn who funded them, so a partisan site can arrive dressed as ordinary local news.
There is a broader democratic stakes question too. Elections depend on a shared set of facts — when to vote, where to vote, what counts — and chatbots now mediate those facts for millions of people. If that mediation is systematically noisier than the official sources it replaces, the cost is not just individual confusion but a widening gap between voters who verify and voters who do not. The researchers' advice, repeated across the reporting, is deliberately low-tech: check the official source.
Election Day is Tuesday, November 3, 2026. In the remaining days, voters still casting ballots or planning to vote in person should verify every logistics detail — registration status, mail-in deadlines, early-voting hours and polling place — against their state or county election office website, or through Vote.gov, rather than relying on a chatbot answer alone. Early voting continues in many states through the weekend, and deadlines that have already passed cannot be recovered by a corrected chatbot answer later.
Expect further studies and company responses after the election, as researchers compare chatbot answers given during the campaign with the official record, and as the companies publish their own post-election assessments. The pattern from the Talarico case — initial repetition, then correction once fact-checkers publish — suggests the tools will keep improving after claims are debunked, but the Forum AI error rate will remain the benchmark against which those improvements are judged. If the next audit still finds roughly four in ten election answers wrong, the case for treating chatbots as a starting point rather than an authority will only strengthen.
For publishers, election officials and the AI companies, the open question is whether the fix is technical, editorial or regulatory: better grounding in official election data, clearer uncertainty signals in answers, or rules that require election information to be sourced from official offices. None of those fixes will be in place before Tuesday. Until they are, the safest advice in this story is also the simplest: ask the chatbot if you like, but confirm it with the people who run the election.
Forum AI found almost 40 percent of chatbot answers to election-related questions contained at least one factual error, across eight models tested ahead of the November 3 midterms.
Logistics errors were more common than candidate errors in the Forum AI study — wrong deadlines, procedures and practical voting details, rather than opinions about candidates.
Google's Gemini gave November 3 as Pennsylvania's mail-in ballot request deadline; the actual deadline was a week earlier. A voter relying on that answer would have missed the window.
Yes, when prompted by NewsGuard, Google's AI Overview, Gemini and Microsoft's Copilot repeated an unfounded allegation about the Texas Senate candidate that AFP fact-checkers had debunked as a likely AI-altered clip bearing hallmarks of a Russian disinformation campaign. Later AFP tests found the tools citing professional fact-checkers instead.
State and county election office websites, and Vote.gov, are the authoritative sources for registration, deadlines, polling places and ballot rules. Researchers advise confirming any chatbot answer about voting logistics against those official sources.
Related coverage: AI scraping opt-out: Google, OpenAI, Microsoft and Anthropic back one no-scrape signal; US midterm elections 2026 coverage; Technology news.
Sources: Wisevoter / Kuwait Times (AFP) reporting on the Forum AI study, November 1, 2026; NewsGuard audits, September 2026; Washington State University, October 28, 2026; AFP fact-checking, late October 2026; company statements from Google and Microsoft as reported by AFP.
Leave a Comment
| Today's Date:
