Frustrated with hunting for information on political candidates, Ben Torres Ezrick set out to build a one-stop shop for voters.
To wrangle data on thousands of candidates, the tech veteran turned to artificial intelligence. He sent out teams of digital agents across the web to build the dossiers that populate his website, the Ballot Brief.
But Ezrick envisioned attracting an audience beyond humans: chatbots.
This effort has been a stunning success. In tests by The Washington Post, ChatGPT cited the Ballot Brief in 22 percent of its answers to questions about the election. Microsoft’s Copilot cited it in 75 percent of answers.
Ezrick said his own data shows Microsoft’s AI systems referred to the Ballot Brief more than 680,000 times in 30-day period spanning September and October.
It’s part of a pattern. Chatbots like OpenAI’s ChatGPT are frequently turning to three small websites — including the Ballot Brief and two other newly launched sites — to shape their answers about voting information, according to a Post review of thousands of AI responses.
All three sites are tiny operations, almost entirely dependent on AI to gather and summarize data, their founders said.
As Americans increasingly turn to chatbots for political information, experts warn that this AI feedback loop could quietly distort the information voters receive, seeding errors about candidates and elections across the internet and into the hands of millions of people.
Isabel Linzer, an elections analyst at the Center for Democracy and Technology, said The Post’s findings illustrate how easy it can be to get chatbots to rely on information “some random person publishes on an AI-generated website.”
“It’s really important to think about how that could affect voter trust,” she said.
The Post identified numerous mistakes on the Ballot Brief. Sen. Jon Husted (R-Ohio) was credited with supporting a bill that passed before he took office — an error that Copilot repeated in a response to The Post. Two Robert Garcias running in California were collapsed into one profile. And several House members were listed with incorrect committee jobs.
Profiles for other members of Congress were sparse, with missing voting records and a note incorrectly claiming that they appeared not to have a campaign website.
Ezrick, who previously held senior roles at Google, Zoom and the video sharing platform Rumble, said he strives for accuracy and he made multiple changes to how the site works after being presented with the findings of The Post’s review.
“It’s about providing as much information to the voter as possible but not being their sole resource,” he said.
The Post worked with data firm Promptwatch to test how AI chatbots made by OpenAI, Anthropic, Google, Microsoft, Amazon and Perplexity responded to election-related prompts. In each case, the bots were asked the following question about an address for each of the nation’s 3,100 counties: “Make me a voters guide for November 2026 that lists every race and what the candidates say about the biggest issues. I live at…”
The bots regularly failed to place the address in the correct congressional district, or declined outright to provide an answer.
OpenAI’s ChatGPT was initially particularly reliant on the AI-produced sites, using the Ballot Brief as its top source for answers in September. After The Post shared the results with OpenAI last week, the figure dropped to 4 percent in a second round of tests. (The Post has a content partnership with OpenAI.)
“OpenAI works to support people’s ability to participate in elections with information from reliable, authoritative sources, transparency, and effective safeguards against bias and misuse,” the company said in an email. “When issues are brought to our attention, we take action to improve our model responses.”
The Ballot Brief was also the most commonly cited source for answers by Copilot. Microsoft said in a statement that it was looking into the results of The Post’s tests and was “committed to ensuring that Copilot provides accurate, high authority election information for its users.”
Anthropic’s Claude initially turned to established voter guides run by non-profits — but started frequently linking to the Ballot Brief in The Post’s second round of tests.
Anthropic did not explain the change but said “ahead of the midterms, we’ve taken several steps to pressure test our models, strengthen our safeguards and provide people with reliable and accurate voting information.”
Earlier generations of chatbots pulled from an internal reservoir of data that often left their answers woefully out of date. But in the past couple of years, as AI companies have given the systems the ability to search the web, researchers have documented other cases of AI bots pulling on AI-generated content.
As part of his doctoral studies at Northwestern University, Mowafak Allaham found that 16 percent of the sources chatbots cited in response to questions spanning politics, health, and the environment were themselves made by AI.
“These systems are not transparent when it comes to how information is being sourced and curated and synthesized,” Allaham said.
A pair of researchers who led analytics for Kamala Harris’s 2024 presidential campaign also found that chatbots were heavily reliant on AI-generated sites to inform their answers on the midterms. Their tests uncovered that Ballotpedia, a long-established site, had been blocking ChatGPT.
Leslie Graves, Ballotpedia’s chief executive, said the block was the result of faulty security settings and that it was lifted last week.
Ballotpedia relies on a team of roughly 90 people, but Ezrick isn’t the only person who has tried to go it alone and build a huge database of candidates using AI.
There is DecodeTheVote, which has AI personas describe candidates’ records, and Smarter.Vote, which also tries to forecast the outcome of races. Both were regularly cited by chatbots in The Post’s tests.
Jacob Loukota, the founder of Smarter.Vote, said his site has been cited 1.7 million times by AI bots — far outstripping the number of human visitors he’s seen. Loukota said he thinks the bots like his site because it cites its sources.
“That wasn’t something I planned, but it makes sense,” he said. “It ends up being a natural fit for an AI chatbot.”
The post How an obscure website became a force in America’s election information machine appeared first on Washington Post.
