AI Assistants Are Pointing Voters to Tiny AI-Written Election Sites

Ask a chatbot for a voter guide at your home address and you usually get a tidy list of races and candidate positions. What you rarely see is the plumbing behind that list — and, according to a test reported by The Washington Post, a surprising share of it now comes from brand-new, lightly staffed websites whose candidate files were compiled largely by AI itself.

That makes this more than a campaign-season story. It is a preview of a machine-to-machine web, where an AI system can be the biggest visitor to a page, where one unverified paragraph can travel into dozens of assistants at once, and where the person reading the answer is left with text that sounds far better sourced than it is.

The Citation Numbers That Triggered the Alarm

In that reporting, ChatGPT cited the site Ballot Brief in 22% of its election-related answers. Microsoft's Copilot did so in 75% of them. The trouble was that Ballot Brief's candidate profiles contained several inaccuracies, and at least one of those mistakes was handed straight back to the reporter who asked the question, in Copilot's own words.

As more voters in the United States turn to assistants for ballot information, the test raises a plain question: if a chatbot leans on an AI-written site, how fast does an error reach the person about to vote?

How One Small Site Became a Favorite Source

Ballot Brief grew out of frustration rather than ideology. Ezrick, who previously held senior roles at Google, Zoom and the video platform Rumble, watched voters bounce between many pages to learn about a single candidate and decided to build a one-stop destination for election information.

Filling it was the harder part. To cover thousands of candidates, he let AI do much of the gathering: he dispatched groups of digital agents to comb the web, then placed their summaries on Ballot Brief for anyone to read.

He noticed quickly that the traffic was not only human. According to data he holds, during a 30-day window that spanned September and October, Microsoft's AI systems referenced Ballot Brief more than 680,000 times.

The wider investigation found that Ballot Brief is one example among several. It and two other recently launched sites — three sites in total — appear often in the source lists assistants hand back for voting questions. All three founders describe small teams and near-total dependence on AI to collect and summarize data.

Isabel Linzer, an elections analyst at the Center for Democracy & Technology, reads those results as a structural problem: information one person publishes on an AI-built site can be adopted by a chatbot with very little friction, and that dynamic can shape how much voters trust the information they receive.

The Mistakes That Leaked Into Answers

Fact-checking by The Washington Post turned up multiple errors on Ballot Brief.

One profile stated that Jon Husted, a Republican senator from Ohio, had backed a particular bill — legislation that had already passed before he took office. Copilot repeated that claim when it answered the reporter's question.

Elsewhere, two California candidates who both go by Robert Garcia were merged into a single profile, and several members of the House of Representatives were listed with the wrong committee assignments. Some records were thin and missing voting history, and the site wrongly indicated that certain legislators appeared to have no campaign website.

Ezrick says he has been working to keep the information accurate, and that after The Washington Post walked him through its findings he made several changes to how the site operates. He adds that the site is meant to give voters as much information as possible, but should not become their only source.

What the Test Actually Asked

The investigation was run with the data company Promptwatch and covered products from OpenAI, Anthropic, Google, Microsoft, Amazon and Perplexity. Researchers picked one address in each of roughly 3,100 counties across the United States and asked every assistant the same thing: produce a voter guide for November 2026 for that address, naming all the races and where candidates stand on major issues.

Performance was uneven. Assistants frequently failed to work out the correct congressional district for a given address, and sometimes declined to answer at all.

One caveat the reporting stresses: the 22% and 75% figures belong to this specific prompt and test environment. They do not mean ChatGPT or Copilot will cite Ballot Brief at those rates across all election questions.

How Each Company Responded

Assistant	What the test showed	What followed
ChatGPT (OpenAI)	Cited Ballot Brief in 22% of election answers; in the first round run in September it was the site ChatGPT used most for this question type	After The Washington Post shared its results last week, the second round showed citations falling to 4%. OpenAI says it wants to help people take part in elections through reliable, authoritative sources plus measures against bias and misuse, and that it acts when problems are pointed out. The Washington Post disclosed in its reporting that it has a content partnership with OpenAI.
Copilot (Microsoft)	Cited Ballot Brief in 75% of answers and repeated the inaccurate Jon Husted claim	Following publication, Microsoft said it adjusted its tools to prioritize authoritative sources and official state election websites, and that it will keep testing and improving election-related answers.
Claude (Anthropic)	Relied mainly on mature voter guides run by nonprofit organizations in the first round, then began linking Ballot Brief frequently in the second round	Anthropic did not explain the change, but said it had tested its model before the midterm elections, strengthened guardrails and works to provide reliable, accurate voting information.
Google, Amazon and Perplexity products	Included in the same six-company test; across the group, assistants often misidentified the congressional district for an address or refused to respond	No product-specific remedy was reported in the findings described.

The Citation Loop: AI Reading AI

Live web access is what makes an assistant feel current, and it is also the doorway through which machine-written material re-enters machine output.

Mowafak Allaham, a doctoral researcher at Northwestern University, found that on questions spanning politics, health and the environment, 16% of the sources chatbots cited were themselves AI-generated. How these systems locate, filter and combine information, he notes, is not transparent.

Two researchers who previously handled data analytics for Kamala Harris's 2024 presidential campaign found comparable dependence on AI-generated websites when assistants answer midterm election questions. Their testing surfaced one unusual detail: Ballotpedia, a site that has long published election information, had been blocking ChatGPT.

Ballotpedia chief executive Leslie Graves said that block was the result of an incorrect security setting and was lifted last week. Her organization runs a team of about 90 people — exactly the kind of staffing Ezrick and others hope to bypass, using AI to build enormous candidate databases at small scale.

Ballot Brief was not the only newcomer flagged. The testing also showed assistants frequently citing DecodeTheVote, which uses AI personas to present candidate backgrounds, and Smarter.Vote, which additionally tries to forecast election results.

Jacob Loukota, the founder of Smarter.Vote, says his site has been cited by AI bots 1.7 million times, far outstripping visits from people. His guess is that assistants favor it because it labels its sources. Loukota adds that this outcome was not part of his original plan, but he finds it understandable, since the site is shaped in a way chatbots can use easily.

Checklist: Six Things to Verify Before You Trust an AI Voter Guide

• Open the source, not only the summary. If an assistant names a site, load that page and see whether it says who wrote the information.
• Look for a named organization and real staff behind it. A site run by a very small team powered by agents deserves different treatment than a newsroom or an official election office.
• Verify the district, not just the candidate. Address-to-district mapping is the step the assistants most often got wrong.
• Check dates, deadlines and measures against official state election websites and long-established voter guides.
• Watch for merged identities. Two candidates sharing one name — the Robert Garcia case above — is a recurring failure pattern.
• Note when you asked. Model behavior can shift between test rounds, as the drop to 4% showed.

Common Misconceptions About AI Election Answers

• A cited source is a checked source. Citation records retrieval, not verification.
• Small and new means it is trying to mislead. Several of the sites in the reporting were built to help voters; the problem was how their output was assembled.
• One fix cures one answer. The inaccurate Jon Husted claim existed on the site and then existed again inside an assistant's reply.
• Lower citation rates mean the risk is gone. The 22% and 75% figures were tied to a single prompt, and the loop of AI reading AI remains.
• Web search guarantees accuracy. It guarantees exposure to whatever is published online, including text produced by other models.

FAQ

Why would an assistant lean on a tiny site at all?

These sites are structured the way retrieval systems like: one clean page per candidate, easy to match to a question. Loukota points to the same reason for Smarter.Vote — it states its sources, which makes it attractive to reference.

Are 22% and 75% meant to describe every political answer?

No. Both numbers come from one prompt, in one test environment, across addresses in roughly 3,100 counties. They describe how the assistants behaved on that task, not their overall reliability.

Was Ballot Brief built to spread misinformation?

The reporting describes the opposite intent: Ezrick created it to spare voters from hopping between pages. Errors were found in its profiles, and he says he has adjusted the site and continues to work on accuracy.

Is the bigger issue about elections or about AI generally?

Elections are where it shows up earliest, because the answers are specific and testable. Northwestern University researcher Mowafak Allaham found the same pattern, at 16%, across politics, health and environmental questions.

What can a reader actually do about it?

Treat an assistant as a starting point rather than a final answer: confirm the jurisdiction, open the cited page, and rely on official state election websites for anything that affects your ballot.

Action Advice for Everyday Readers

If you plan to ask a chatbot anything about your ballot, do it in three steps rather than one. First, ask for sources by name and write them down. Second, spend one minute confirming the district, office and election date against an official state page or an established nonprofit voter guide. Third, treat anything unusual — a missing campaign website, a strange committee listing, two candidates collapsed into one biography — as a prompt to check before sharing.

For people who publish or build products, the same reporting carries a harder lesson. When your content, your profile or your data is machine-readable, a model may already be your largest audience, and any inaccuracy becomes a template that others copy. The practical response is boring but effective: label sources, name a responsible team, correct errors visibly, and assume that what you publish can be quoted by a bot before a human ever reads it.

We will keep tracking how assistants handle civic and consumer information. Readers who spot an AI answer worth checking can reach Tech Daily at hello@techdaily.com. By Sarah Chen, MyTechDailyNews.