AIdataNewsGlobal General Briefing: Why Chatbots Keep Sending Voters to Tiny AI-Built Election Sites

Ask a chatbot for a voter guide and the answer you receive may have been assembled not by election officials or long-established reference desks, but by a handful of brand-new websites whose founders used AI to gather and summarize candidate data. That is the picture emerging from a series of tests published by The Washington Post, which measured how often commercial assistants lean on a small number of such sites when answering questions about United States elections. For readers tracking AI and data governance, the findings raise a structural question: once machines become the main audience for machine-made content, who verifies the facts before they reach voters?

The numbers that frame the story

• ChatGPT cited Ballot Brief in 22% of its election-related answers.
• Microsoft Copilot cited the same site in 75% of them.
• In a 30-day window spanning September and October, Microsoft's AI systems referenced Ballot Brief more than 680,000 times.
• The test covered roughly 3,100 counties across the United States, using one address per county.
• Three new sites, Ballot Brief among them, appeared repeatedly in the voting sources chatbots returned.
• 16% of the sources chatbots cited on politics, health and environment questions were themselves AI-generated, according to a study by a PhD researcher at Northwestern University.
• ChatGPT's citation rate for Ballot Brief fell to 4% in a second round, after The Washington Post shared its results with OpenAI last week.
• Smarter.Vote reports its pages have been cited by AI bots 1.7 million times, far exceeding traffic from human readers.

Citation patterns, assistant by assistant

Assistant	What the test recorded	What the company said or did
ChatGPT (OpenAI)	Ballot Brief was its most-cited source for this question in the first, September round, at 22%.	OpenAI says it wants to help people take part in elections through reliable, authoritative sources and measures against bias and misuse, and that it acts when problems are identified. The rate dropped to 4% in the second round.
Copilot (Microsoft)	75% of election answers cited Ballot Brief, and the assistant repeated one of the site's factual errors to a questioner.	Microsoft said after publication that it adjusted the tool to favor authoritative sources and official state election websites, and will keep testing and improving election-related answers.
Claude (Anthropic)	Relied mainly on established voter guides run by nonprofits in round one, then began linking to Ballot Brief frequently in round two.	Anthropic did not explain the change. It said it has tested its model ahead of the midterms, strengthened safeguards and works to provide reliable, accurate voting information.
Across the test set	Products from OpenAI, Anthropic, Google, Microsoft, Amazon and Perplexity were examined.	Beyond Ballot Brief, chatbots frequently cited DecodeTheVote and Smarter.Vote.

How the test was built

1. The Washington Post worked with the data company Promptwatch.
2. Assistants from OpenAI, Anthropic, Google, Microsoft, Amazon and Perplexity were put through the exercise.
3. Researchers selected one address in each of roughly 3,100 U.S. counties.
4. Each assistant received the same prompt: produce a November 2026 voter guide for that address, listing every race and the candidates' positions on major issues.
5. Outcome: the assistants often failed to identify the congressional district an address belongs to, and sometimes declined to answer altogether.
6. Scope note: the percentages describe this specific prompt and test environment, and do not imply that ChatGPT or Copilot cite Ballot Brief at those rates for every election question.

Ballot Brief: one small team, thousands of candidate files

Ezrick, a former senior executive at Google, Zoom and the video platform Rumble, watched voters bounce between many sites just to learn who is on the ballot, and decided to build a one-stop election information website. Putting together files on thousands of candidates was the point where he turned to the technology: he dispatched multiple groups of digital agents to search for information online, then loaded their summaries into Ballot Brief for people to consult.

Its visitors, he says, are not only human. According to data he holds, Microsoft's AI systems cited Ballot Brief more than 680,000 times over a 30-day period spanning September and October. The Washington Post found his site to be one example of a wider pattern: three new sites, Ballot Brief included, turn up often among the voting sources chatbots provide. Founders of all three said their teams are small and rely almost entirely on AI to collect and summarize data.

The errors that reached voters

Checking Ballot Brief against the record turned up several problems:

• One profile stated that Jon Husted, a Republican senator from Ohio, had backed a bill that had already passed before he took office. Copilot repeated this error when answering a question from The Washington Post.
• Two California candidates who share the name Robert Garcia were merged into a single profile.
• Committee assignments for several members of the House of Representatives were listed incorrectly.
• Some lawmakers' files were thin and lacked voting records, and the site wrongly indicated that they appeared to have no campaign website.

Ezrick said he has been working to keep the information accurate, and that once The Washington Post showed him the findings he made multiple adjustments to the way the site operates. He added that the site is meant to give voters as much information as possible, but should not become their only source.

When AI starts citing what AI wrote

Web-connected search lets assistants reach information newer than their training data, but it also exposes them to the growing volume of AI-generated material online. Mowafak Allaham, a PhD researcher at Northwestern University, found that on questions in politics, health and the environment, 16% of the sources cited by chatbots were themselves AI-generated. How these systems locate, filter and combine information is not transparent, he says.

Two researchers who previously handled data analysis for Kamala Harris's 2024 presidential campaign reached a similar conclusion: assistants answering midterm election questions leaned heavily on AI-generated websites. Their testing also found that Ballotpedia, a long-established election information site, had been blocking ChatGPT. Ballotpedia CEO Leslie Graves said the block was caused by a misconfigured safety setting and that the restriction was lifted last week. Her organization runs a team of about 90 people, the kind of scale Ezrick and others hope AI can compress while still building very large candidate databases.

The other two sites in the chatbots' answer set take different routes: DecodeTheVote uses AI personas to present candidates' backgrounds, while Smarter.Vote also attempts to forecast election results. Smarter.Vote founder Jacob Loukota said his site has been cited by AI bots 1.7 million times, far surpassing visits from humans. He believes assistants may favor it because the site marks its sources, and said the result was never part of his plan but is understandable, since the site's format suits AI chatbots.

Stakeholder reactions in brief

• Isabel Linzer, an election analyst at the Center for Democracy & Technology, reads the findings as evidence that information one person publishes on an AI-generated website can be adopted by a chatbot with little friction, a dynamic she says may affect voters' trust in election information.
• OpenAI frames its election work around reliable, authoritative sources plus defenses against bias and misuse, and says it acts to improve model answers when problems are reported.
• Microsoft has re-prioritized authoritative sources and official state election websites in Copilot.
• Anthropic says its model was tested before the midterms and its safeguards strengthened.
• Ballotpedia has removed the block on ChatGPT; Ballot Brief made several operational changes.

One disclosure recorded in the reporting: The Washington Post has a content partnership with OpenAI.

FAQ: what readers ask about the findings

Does 22% mean ChatGPT cites Ballot Brief in a fifth of all election questions?

No. The 22% and 75% figures come from one prompt, a request for a November 2026 voter guide tied to a specific address, under one set of test conditions.

Did these small sites set out to supply chatbots?

Not as a plan. Ezrick built Ballot Brief because voters were hopping between sites; Loukota said the volume of bot citations was not what he expected, only that he finds it understandable given the site's format.

What changed once the results were shared?

ChatGPT's citation of Ballot Brief fell to 4% in the second round after The Washington Post informed OpenAI last week; Microsoft adjusted Copilot to favor authoritative and official state election sources; Ballotpedia lifted its block on ChatGPT.

Why does this matter outside U.S. elections?

It is a working example of an AI-cites-AI loop, and the reason the Northwestern researcher's point about opacity is a governance issue rather than a technical footnote: the same retrieval path carries information in politics, health and the environment, where 16% of cited sources were themselves AI-generated.

What to watch next

The citation mix moved within weeks of a single set of tests, and at least one source's share fell from 22% to 4% after the developer was notified. That responsiveness is one half of the story; the other is that the assistants still depend on archives assembled by very small teams using AI, including the three new sites in this test. The next checkpoint is November 2026, when voters across some 3,100 counties may ask a chatbot a local question that only a handful of machine-written sites, alongside official state election pages and established guides such as Ballotpedia, are trying to answer. How those sources settle into a citation order is what this data leaves open.