Nearly 40% of AI chatbot election answers contain a factual error
A Forum AI study of eight chatbots found that almost 40% of their answers to election-related questions contained at least one factual error, according to AFP's report of 1 November 2026. Questions about how to vote produced more errors than questions about candidates. One example involved Google's Gemini 3.6 Flash, which gave 3 November as the Pennsylvania deadline to request a mail-in ballot. The official application deadline was 27 October. For businesses that let AI tools answer staff or customer questions about dates and rules, the finding is a direct warning: a fluent answer can still be wrong on the one detail that matters.
In short
- Forum AI found nearly 40% of answers from eight chatbots had at least one factual error on election questions.
- Voting-logistics questions went wrong more often than candidate questions.
- The Pennsylvania mail-in ballot error was a week off the official application deadline of 27 October 2026.
- Any business that lets AI answer deadline or rule questions needs an official source it controls and a regular sample check.
- The May 2026 Forum AI study used a different test and should not be compared directly with the November figure.
What happened
Forum AI, a start-up that evaluates AI models, tested eight chatbots on election questions. According to AFP's report published on 1 November 2026, nearly 40% of the answers contained at least one factual error. The study's researchers said chatbots got voting-logistics questions wrong more often than questions about candidates. The test covered questions about the US midterm elections, which take place on 3 November 2026.
The best-known example concerns Pennsylvania. When Google's Gemini 3.6 Flash was asked about voting in the state, it gave 3 November as the deadline to request a mail-in ballot. The actual deadline was a week earlier. Forum AI wrote that a voter who followed the answer would have asked for a ballot too late to use it. Pennsylvania's Department of State had set the application deadline for 5 p.m. on 27 October 2026, and completed mail ballots must reach county offices by 8 p.m. on Election Day. AFP reported that Google said it has strict quality protections for election topics, but did not explain the Pennsylvania error.
Other errors in the study were of a different kind. Forum AI said chatbots attributed incorrect statements to candidates, officials and observers hundreds of times. In one case, Anthropic's Claude described North Carolina Representative Don Davis as a former school superintendent, confusing him with another person. Forum AI also said the chatbots often cited weak sources, including Reddit threads.
The figure is not a repeat of an earlier result. Forum AI published a May 2026 study that found nearly 36% of election answers contained at least one factual error. That test was run differently, so the two numbers should not be read as a trend line. Microsoft told AFP that Copilot had updated its answer once more information became available, and that it later identified a disputed claim as fabricated.
Why it matters for businesses that automate work
Many businesses now put AI in front of routine questions. Staff ask an internal assistant when a payroll run closes. Customers ask a chatbot on a website when an invoice is due or when a return must be filed. The Forum AI study is not about payroll or invoices, but it tests the same weakness: a tool that answers in a confident tone can fail on a date, a threshold or a procedure.
The error pattern matters more than the headline rate. Procedural errors are the ones that cost people something. A wrong date for a ballot is a missed vote. A wrong cut-off for a supplier payment is a late fee, a broken contract term or a missed filing. Automated answers about deadlines therefore need a different check from answers about general knowledge.
What changes in practice, at any size
A five-person office and a firm with thousands of staff face the same basic choice. Either the AI answers from a source the owner controls, or it answers from whatever it learned during training and the web. The first option is safer for deadlines and rules.
| Type of question | Safer setup | What a person still checks |
|---|---|---|
| Deadlines and filing dates | Answer only from a dated calendar the owner maintains, with a link to the official page | The calendar, after each official change |
| Rules and thresholds | Answer only from the official text or the regulator's page, quoted by link | Any answer the tool cannot match to a page |
| Customer or staff questions about the business itself | Answer from the company's own policy documents, dated and versioned | Refund, exception and complaint decisions |
| Opinion or open-ended questions | Route to a person | Everything |
Three steps follow from the study. First, list the questions where a wrong answer has a deadline or cost attached. Second, decide which official page is the source for each one and write that link into the tool's instructions. Third, sample the answers every week and compare them with the source. A chatbot that changes its answers after a correction, as Microsoft described for Copilot, still needs someone to confirm the new answer.
Automation has limits here. A tool can retrieve a page and draft a reply, but it cannot know that a rule changed unless someone updates the source. It cannot judge whether a question falls outside its material. These decisions stay with a person. Where an error would cause a missed filing or a legal problem, the question should go to a qualified person, not an automated answer.
What to watch next
Three things will show whether the problem is narrowing. The first is whether Forum AI publishes the model-level results behind the November figure. Forum AI did not name the model in one report, while AFP named Gemini 3.6 Flash, so the attribution rests on AFP's reporting. The second is whether the chatbot makers correct the specific errors. Google and Microsoft have both said their systems updated after the tests. Anthropic did not respond to AFP's request for comment. The third is whether any regulator in the United States sets rules for AI answers on voting. AFP reported in its coverage that Congress has not passed laws regulating AI in politics, so for now the standard is set by the companies and by the officials who run elections.
Sources
How AiStaffo would automate this
AiStaffo would start by listing the questions a business gets most often that carry a date or a rule, such as filing dates, invoice terms or supplier cut-offs. The AI worker would be connected to a deadline calendar and a policy folder that the owner controls, with each entry linked to its official page. It answers only from those sources, shows the link with every reply, and sends anything it cannot match to a named person. The owner or the finance lead still updates the calendar when a rule changes, reviews a weekly sample of answers, and decides every exception. To see which of your own questions carry a deadline, book a free automation audit.
Questions people ask
What did the Forum AI study find about chatbot election answers?
Is this the same as the earlier Forum AI study?
What was the Pennsylvania deadline that the chatbot got wrong?
Does this matter for businesses that use AI for customer or staff questions?
What should a business do first if it already uses a chatbot for rule questions?
Book a free automation audit
Thirty minutes. We look at one process you run every week and tell you exactly what an AI worker would take off your desk, and what it would not.
