What did the latest audit of AI election answers find?
Americans vote in the midterm elections on November 3, 2026. AFP reported on November 1 that this is the first federal election in which millions of voters actively rely on AI tools for political information. The same report described a pattern that researchers say could affect how people vote.
The clearest number comes from Forum AI, a startup that evaluates model behaviour. Its election benchmark covers eight leading models and 26,398 answers to questions about the 2026 races. Forum found that 10,205 answers, or 38.7%, contained at least one false verifiable claim.
The benchmark's own headline puts it bluntly: almost 40% of election answers contained at least one factual error. The company tested models from Anthropic, OpenAI, Google, xAI, Alibaba, DeepSeek, Zhipu AI and Moonshot AI.
The stakes are broad because every seat in the House is on the ballot this year. Forum's prompts covered candidate races, ballot logistics and kitchen-table economic issues. All of those answers can reach voters who never open a news site.
Errors clustered on the practical questions voters ask closest to polling day. Forum says models handled questions about how to vote 13% worse than questions about candidates. They also did worse on logistics than on broader policy topics.
One example involved Pennsylvania's mail-in ballot deadline. Forum asked Gemini 3.6 Flash about the state's 2026 ballot and received an answer that the request deadline was November 3. That is Election Day itself. The real deadline to apply was 5 p.m. on October 27.
Forum's judge wrote that a voter who followed the answer would have asked for a ballot too late to use it. The benchmark also recorded cases where models put the wrong words in candidates' mouths. It counted those quote errors hundreds of times across the sample.
Which chatbots were most accurate on election questions?
Accuracy varied widely between models, even within the same benchmark and prompt set. Forum reports the share of answers that contained at least one false claim for each model in its sample. Lower is better.
The spread is large enough to matter for anyone choosing a tool. GPT-5.6 (Terra) from OpenAI had false claims in 11.8% of its answers. DeepSeek V4 Pro had them in 60.2%.
Claude Opus 5 from Anthropic landed at 51.5%, the second-worst result. Google's Gemini 3.6 Flash recorded 39.1%, almost exactly the overall average.
The middle of the field was tight. Grok 4.5 from xAI recorded 32.5%, GLM-5.2 from Zhipu AI 35.0%, Alibaba's Qwen 3.8 Max 39.4% and Moonshot's Kimi K3 39.9%.
Forum also scored whether answers introduced unrequested political bias. About 7% of election answers did. That is lower than the error rate, but it matters when a voter asks for a neutral summary of two candidates and receives a verdict instead.
- •GPT-5.6 (Terra), OpenAI: 11.8% of answers with a false claim
- •Grok 4.5, xAI: 32.5%
- •GLM-5.2, Zhipu AI: 35.0%
- •Gemini 3.6 Flash, Google: 39.1%
- •Qwen 3.8 Max, Alibaba: 39.4%
- •Kimi K3, Moonshot AI: 39.9%
- •Claude Opus 5, Anthropic: 51.5%
- •DeepSeek V4 Pro, DeepSeek: 60.2%
Where do chatbot election errors hit voters hardest?
A separate audit by the Institute for Strategic Dialogue, published in September 2026, looked at state-level rules rather than national races. ISD tested six models with 400 prompts each across ten states. Its total was 2,400 prompts and responses.
ISD found that 29% of responses to generic English prompts were incomplete, inaccurate or outdated. It classified 12% as having the potential to meaningfully mislead users and affect their ability to vote.
The basic facts were not immune. Meta's Muse Spark identified Election Day as November 4, 2026, in two responses. The correct date is November 3.
Language changed the results. Accuracy fell 16% overall when the same prompts were asked in Spanish. Four models failed to give accurate and complete answers to more than 40% of Spanish prompts.
The Washington Post ran a third test with a different method and reached a similar conclusion. Reporters asked six popular chatbots to build voter guides for a random address in every US county. They then analysed more than 3,000 guides.
The chatbots failed to say which congressional district an address was in 48% of the time. When they did name a district, they got it wrong 22% of the time. Every voter in those counties had a House race on the ballot.
The pattern follows the way the models find information. Election rules differ by state and often by county, and they change between cycles. ISD describes those thin spots as data voids. Adversaries and well-meaning publishers alike can fill them with pages that look authoritative.
Citation habits reflect that gap. ISD found that answers to generic voter questions leaned on government sources. Answers to adversarial prompts leaned on news media instead. Only Meta's Muse Spark showed a clear tendency to cite social media.
ISD also tested whether models would accept false claims introduced by a user. All six refuted them at high rates, and none affirmed a false claim more than 4% of the time. The weak point was hedging. Muse Spark and Gemini gave ambiguous answers to 22% and 18% of misleading claims.
A wrong district is more damaging than a wrong talking point. Voters who are handed the wrong House race may study candidates they cannot vote for and miss the ones they can. That error is hardest to spot, because the answer still looks complete.
Why chatbots cite AI-written voter sites
The Post's October 30 follow-up traced part of the problem to sourcing. Its tests found that Microsoft's Copilot cited a site called the Ballot Brief in 75% of voter-guide responses. ChatGPT cited it in 22%.
The Ballot Brief is not a neutral, long-established reference work. Its own about page says the project uses AI to research and write election information. It also warns that errors are possible and asks readers to verify important details with official sources.
The site is open about its limits. It says a ZIP code search is less precise than a full street address and may omit state legislative and local races. It also says automated checks run before each update, but no one reads every profile before it goes live.
The Post reported that it found multiple cases where the site translated ZIP codes to the wrong congressional districts. Chatbots then repeated those errors, according to the report. That turns one small mapping mistake into repeated answers at scale.
Scale is the point. Ben Torres Ezrick, the site's founder, told the Post that Microsoft's AI systems referred to the Ballot Brief more than 680,000 times in a 30-day period spanning September. The site describes Ezrick as a former Google and Zoom executive based in Miami Beach.
The Post framed the Ballot Brief as part of a new crop of influential AI-created election sites. Those pages are built to be found by search engines and answer engines. Retrieval systems reward that visibility, not necessarily the accuracy of the underlying data.
The site is also part of a network. Its pages link to two sister projects, one on community budgets and one on homebuying assistance. Each uses the same plain, structured format, which helps explain why answer engines keep finding them.
For a reader, a citation looks like a source. The page may be the only place with a full candidate list for a down-ballot race, even when its district data is wrong. Corrections arrive only when someone notices and reports them, the site says.
What do Google, Microsoft and Anthropic say?
Company responses have focused on safeguards and on how quickly systems update. Anthropic did not respond to AFP's request for comment on the Forum AI findings. The report included an error involving Claude.
Google said it is focused on helping voters find reliable information across its platforms. It pointed to strict quality and policy protections for election topics across its AI products.
Microsoft answered questions about a different test. NewsGuard, a disinformation watchdog, found that ten AI tools repeated a likely AI-altered clip about Texas Senate candidate James Talarico. Google's AI Overview, Gemini and Copilot all repeated the false claim, according to AFP.
Microsoft described that episode as a rapidly developing situation with limited reliable information at the time of the initial query. It said Copilot later reflected updated context and identified the claim as fabricated.
Google said the issue was addressed as its systems quickly updated to surface the latest reporting. AFP's own later test found the tools citing professional fact-checkers, which shows how much answers can shift within days.
The statements address specific episodes rather than the audits as a whole. They argue that systems improve as information changes, not that the earlier answers were correct.
How reliable are these election audits?
The three main studies used different models, prompts and scoring, so their percentages are not directly comparable. Forum counted 26,398 answers, ISD 2,400 and the Post more than 3,000 voter guides. Each tested a snapshot of models that change frequently.
Forum's benchmark is dated August 2026. ISD's report was published in September. The Post's tests ran in October. Some of the model versions named in the studies may already have been updated.
Scoring also involves judgment. Forum uses calibrated judges to decide whether a response contains a false verifiable claim. ISD coded responses for accuracy, timeliness and sourcing. A single wrong phone number counts as an error in both approaches, even if the rest of an answer is sound.
ISD is candid about its own constraints. Researchers collected the data in mid-June 2026 and ran each prompt once, so the results are a snapshot rather than a stable performance gap. Five of the six models were queried through developer APIs, which may not match the consumer apps voters use.
The ten states were chosen partly for recent rule changes, active litigation or a history of election administration controversies. ISD says that makes its findings a picture of high-risk contexts, not a national average. Meta's model was also tested manually, because it lacks comparable API access.
Forum publishes other measures that put the accuracy number in context. About 42% of answers that used well-rated sources still contained a false claim. Roughly 35% of all answers drew on sources it scored below 75 out of 100, and about 53% cited at least one commercial source.
There is counter-evidence too. AFP reported that research by the Brennan Center for Justice found leading chatbots pushed back against common election conspiracy theories. That included claims that voting machines were rigged, even when users challenged the models repeatedly.
The Brennan Center said the chatbots often made mistakes but also listened sympathetically to skeptical users without endorsing conspiracy theories. It suggested they could play a constructive role against election disinformation.
Voters are not fully convinced either. AFP cited a National Conference on Citizenship survey in which two in five Americans said they had used AI tools to search for election information. Only 40% trusted them with logistical details such as when, where and how to vote.
The same survey found that 60% trusted the tools only a little or not at all to be unbiased about candidates.
What should voters do before November 3?
The practical lesson from the audits is that chatbots are a starting point, not an authority. The errors that matter most involve deadlines, eligibility, polling places and district numbers. Those are exactly the details that state and county election offices publish directly.
A few habits reduce the risk. Use a full street address rather than a ZIP code when checking a ballot or district. Ask the chatbot for the source of a specific claim, then open that source and confirm it says what the answer claims.
Treat quotations and candidate biographies with extra care. Forum recorded hundreds of cases where models attributed words to candidates or officials incorrectly. It also found that errors were more common on less competitive races, where reliable coverage is thinner.
Official sample ballots and state election websites remain the reference point. The Ballot Brief makes the same recommendation on its own site. So do the researchers behind the audits, who note that rules differ by state and often by county.
Some errors have already outlived the process they describe. Pennsylvania's deadline to apply for a mail-in ballot passed on October 27, so an answer that still names November 3 sends voters looking for a form that is no longer available.
Several things are worth watching after November 3. One is whether Copilot and other assistants change which sources they cite for election queries. Another is whether AI-written voter sites correct their district mapping at scale, rather than one page at a time.
A third is whether the accuracy gap between models narrows in the next benchmark. The current spread, from 11.8% to 60.2% of answers with a false claim, is far wider than most users would expect from tools marketed for the same task.
For now, the evidence points one way. A confident answer about a ballot deadline can still be wrong, and the cost of believing it falls on the voter.