Abstract:
When American voters ask ChatGPT or Microsoft Copilot for candidate information, the answer they get may come from a small website they have never heard of. The website, called Ballot Brief, collects profiles of thousands of candidates. Its creator, Ben Torres Ezrick, hopes to help voters understand campaign information more easily and also hopes to attract a special type of visitor—AI chatbots.
A test by The Washington Post showed that ChatGPT cited Ballot Brief in
22% of election-related responses, compared with 75% of Microsoft Copilot
. However, there are many errors in the candidate information on this website, one of which was repeated by Copilot to the questioner.As more Americans ask AI for political and election information, the survey begs the question: When chatbots rely on AI-generated websites, how do errors get to voters?
A small website has become a common source of AI
Ezric has held senior positions at Google, Zoom and video platform Rumble. He found that when voters wanted to learn about candidates, they often had to spend a lot of time searching between different websites, so he decided to build a one-stop website that provided election information.
Faced with the information of thousands of candidates, he chose to let AI help. He sent multiple groups of digital agents to collect information online, and then put the compiled content into Ballot Brief for people to review.
Websites attract far more visitors than just humans. Ezrick said that the data he had showed that in the 30 days spanning September and October, Microsoft's AI system cited Ballot Brief more than
680,000 times
.The Washington Post’s investigation found that Ballot Brief was just one of them. It is among three new websites that frequently appear among the sources of voting information provided by chatbots. The founders of the three websites all said that their teams are small and rely almost entirely on AI to collect and summarize data.
Problems arise. Isabel Linzer, an elections analyst at the Center for Democracy and Technology, believes that the survey results show that information posted by a person on an AI-generated website may easily be adopted by a chatbot. That could affect voters' trust in election information, she said.
The mistake was repeated by the AI
The Washington Post found multiple errors in the Ballot Brief.
An information on the website stated that Ohio Republican Senator Jon Husted sponsored a bill. But that bill passed before he took office. Copilot repeated this error in response to questions from The Washington Post.
The website also merged two California candidates with the same name, Robert Garcia, into the same profile, and the committee assignments of several House members were also listed incorrectly.
The information of some congressmen is quite brief, lacking voting records, and the website incorrectly states that they do not seem to have a campaign website.
Ezric said he works hard to ensure the information is accurate. After The Washington Post showed him the findings, he made a number of changes to how the site worked.
The purpose of the website is to provide voters with as much information as possible, but it should not be their only source of information, he said.
A test covering 3,100 counties at one time
To see how chatbots answer election questions, The Washington Post partnered with data firm Promptwatch to test AI products from OpenAI, Anthropic, Google, Microsoft, Amazon and Perplexity.
Researchers targeted approximately
3,100 counties in the United States
, pick an address in each county, ask the chatbot the same question, and ask it to create a November 2026 voter guide based on that address, listing all elections and candidates’ positions on major issues.As a result, chatbots often failed to correctly determine which congressional district an address belonged to, and sometimes simply refused to answer.
The above citation ratios are from this specific issue and test environment and do not mean that ChatGPT or Copilot will quote Ballot Brief at the same ratio in all election questions.
The Washington Post found that during its first round of testing in September, Ballot Brief was ChatGPT’s most frequently cited source when answering such questions. After the newspaper notified OpenAI of the results last week, ChatGPT's share of citations to the site dropped to
4%
in a second round of testing. .OpenAI stated that the company hopes to help people participate in elections through reliable and authoritative sources of information and measures to prevent bias and abuse; when issues are pointed out, the company will take action to improve model answers. The Washington Post disclosed in its report that the newspaper has a content partnership with OpenAI.
The Ballot Brief is also Copilot's most frequently cited source in testing. After the report was published, Microsoft said it had adjusted the tool to prioritize authoritative sources and official state election websites and would continue to test and improve election-related responses.
Anthropic’s Claude mostly cited established voter guides run by nonprofits in the first round of testing, but started linking to Ballot Brief more frequently in the second round of testing. Anthropic did not explain the change, saying only that the company had tested models, strengthened safeguards and worked to provide reliable, accurate voting information ahead of the midterm elections.
When AI starts to quote content written by AI
Internet search allows chatbots to access newer information than their training data, and also gives them greater access to AI-generated content on the Internet.
A study by Mowafak Allaham, a doctoral researcher at Northwestern University, found that **16%** of the sources cited by chatbots on issues in areas such as politics, health, and the environment were themselves generated by AI.
How these systems find, filter and integrate information is not transparent, he said.
Two other researchers who worked on data analysis for Kamala Harris’ 2024 presidential campaign found that chatbots also relied heavily on AI-generated websites when answering midterm election questions. Their testing also found that Ballotpedia, a website that has long provided election information, had previously blocked ChatGPT.
Ballotpedia CEO Leslie Graves said the block was caused by an error in security settings and that the site had lifted the restrictions last week. Ballotpedia has about
90 people
The team, Ezrick and others hope to use AI to build a huge candidate database on a smaller scale.In addition to Ballot Brief, The Washington Post's testing also found that chatbots frequently referenced DecodeTheVote and Smarter.Vote. The former uses AI characters to introduce candidates’ resumes, and the latter also attempts to predict election results.
Smarter.Vote founder Jacob Loukota said his website has been cited by AI bots
1.7 million times
, far exceeding the number of visits by human visitors. He thinks the chatbots like to cite the site, possibly because it identifies the source of the information.Lukota said that this was not his original plan, but he felt that the result was understandable and that the form of the website was very suitable for the use of AI chatbots.
Comments