How an AI-built voter site became a chatbot source
In an October 30 follow-up to its investigation of AI-generated voter guides, The Washington Post reported that Microsoft Copilot cited the Ballot Brief in 75% of tested election responses, while ChatGPT cited it in 22%. The Ballot Brief is a small website that uses AI tools to assemble information about political candidates.
The finding followed the Post's earlier October investigation of more than 3,000 chatbot voter guides. That analysis found that 48% of guides did not identify the voter's congressional district. Among guides that did identify a district, 22% named the wrong one.
The two findings describe different weaknesses in the same information pipeline: chatbots can repeatedly cite inaccurate third-party material, and their resulting voter guides can omit or misstate essential ballot information.
The Post identified a case in which the Ballot Brief incorrectly credited Ohio Sen. Jon Husted with supporting legislation passed before he entered office. Copilot repeated that error in a response to the newspaper. The Post also found that the site merged profiles of two California candidates named Robert Garcia and misstated committee assignments for several House members. Those examples show why a citation is not itself an accuracy check: the chatbot can reproduce errors already present in its source.
The responses on record
The Ballot Brief, at theballotbrief.com, is described by the Post as largely produced by AI tools and run by Ben Torres Ezrick, a former executive at Rumble. The Post found multiple instances of the site translating ZIP codes to the wrong congressional districts, errors that chatbots citing the site then repeated.
Ezrick said his site draws on official government sources and lets users enter an address for more complete candidate information. OpenAI's Nate Evans said ChatGPT's search tools cite a range of sources so people can explore and make their own judgements, and that OpenAI acts to improve responses when issues are raised.
The Post also noted that OpenAI and Anthropic have both announced partnerships intended to give their chatbots trusted election information. Those efforts did not prevent the errors measured in the earlier investigation.
The U.S. Vote Foundation's summary of the Post's work records how concrete some of the errors were. For a Missouri address, ChatGPT listed a US Senate seat that is not on the ballot this year. For a Wyoming address, Claude listed local politicians as candidates for federal office. Some guides reported a Senate race in states that have none this year, even though Senate seats follow a predictable six-year cycle.
Neither response addresses the concentration the October 30 tests measured: one AI-built site supplied three quarters of one chatbot's citations. The tools are supposed to draw on a range of sources. In the Post's tests, Copilot's sourcing narrowed to a single small site whose ZIP-to-district mappings the Post found wrong in multiple instances.
MIT is building a longer-term record
Separately, MIT researchers announced the LLM Election Observatory on October 30. Led by Adam Berinsky, Chara Podimata and Charles Stewart III, the project repeatedly submits election-related questions to popular chatbots and records how responses vary over time, location, political affiliation and simulated user identity.
The project is broader than ballot accuracy. It also examines sourcing, framing and how election information changes as models and their search systems are updated. MIT says its full longitudinal analysis is not yet complete.
For voters, the immediate lesson is narrower: use state and local election offices to confirm district boundaries, candidates and deadlines. A chatbot-generated guide may help identify questions to investigate, but it should not be the final authority on an individual ballot.
Sources
- [1] AI Weekly, “Copilot Cited AI-Built Ballot Brief in 75% of Election Tests”, November 1, 2026 (a summary of The Washington Post's reporting, read first-hand: Copilot cited the Ballot Brief in 75 percent of voter-guide tests and ChatGPT in 22 percent; six chatbots asked for guides for a random address in every US county and more than 3,000 guides analyzed; no congressional district named 48 percent of the time and the wrong one named 22 percent of the time one was named; the Ballot Brief described in the Post's words as largely produced by AI tools and run by a former executive at Rumble, with multiple instances of ZIP codes translated to the wrong congressional districts; and the responses from the site's founder, Ben Torres Ezrick, and OpenAI's Nate Evans)Read source
- [2] Techmeme snapshot, October 31, 2026, 20:00 ET (the Post's headline finding: Copilot cited the AI-built Ballot Brief in 75% of voter questions, ChatGPT in 22%; chatbots relied on three obscure AI-built voter sites in total)Read source
- [3] U.S. Vote Foundation, “AI Chatbots Fail Voters”, read first-hand (the Post's analysis of more than 3,000 AI-generated voter guides from six chatbots, by reporters Kevin Schaul and Ian Duncan; the Missouri address given a US Senate seat that is not on the ballot this year; the Wyoming address where Claude listed local politicians as candidates for federal office; guides reporting a Senate race in states that have none this year, although Senate seats follow a predictable six-year cycle; and the note that OpenAI and Anthropic have both announced partnerships to give their chatbots trusted election information, which did not prevent significant errors)Read source
- [4] MIT SHASS / MIT News Q&A, “Hey AI, who should I vote for this year?”, October 30, 2026, read first-hand (the LLM Election Observatory led by Adam Berinsky, Chara Podimata and Charles Stewart III: a fixed set of election questions re-run through popular chatbots every few days, varying location, party affiliation and identity, with the underlying data published)Read source