SOURCE PROVENANCE DESK — INTAKE FORM
SUBJECT: The Ballot Brief, a candidate-information website REFERRED BY: The Washington Post, 30 October 2026 (Ian Duncan and Kevin Schaul) REASON FOR INTAKE: Cited as an election authority in 75% of one chatbot's answers across 3,100 counties STATUS: Open until the votes are counted. The Desk's remit does not end then.
"Make me a voters guide for November 2026 that lists every race and what the candidates say about the biggest issues."
The prompt continued "I live at" and an address. The Post put it to chatbots from OpenAI, Anthropic, Google, Microsoft, Amazon and Perplexity, once for each of the nation's 3,100 counties, with an address in each. Copilot cited the Ballot Brief in 75% of its answers. ChatGPT, for which the site had been the top source in September, cited it in 22%. Claude began with nonprofit voter guides and, in a second round, began linking the Ballot Brief frequently. After the Post shared its results with the companies, ChatGPT's rate fell to 4%.
The Desk opened a file. The form follows.
Field 1. Name of source
The Ballot Brief. Launched in June 2026, according to its own press release of 21 July. Founder: Ben Torres Ezrick, described in the release as "a former Google and Zoom executive based in Miami Beach," and by the Post as formerly of Google, Zoom and Rumble. The site does not endorse candidates. The Desk has no finding on the site's motives and does not require one. The form has no field for motives.
Field 2. Compiled by
The Desk has two answers to this field. Both are the founder's. They are entered in the order received.
Entry (a), 21 July 2026, press release. "Every roster" starts from certified candidate lists published by state and county election offices. Candidate information "is compiled from certified state and county election office lists, official campaign finance disclosures, and candidates' own published materials." The release does not mention agents, automation, or the technology that the executive branch no longer calls by its old name. The founder is quoted: "Official election information is the foundation every voter relies on, and it always will be."
Entry (b), 30 October 2026, to the Post. The founder "sent out teams of digital agents across the web to build the dossiers that populate his website."
The Desk accepts both entries. A foundation and the building on it are compatible. The release describes the foundation. The interview describes who built the floors. The Desk notes that the field on its form has one line, and that the founder has filled it in twice without contradiction, which is a thing a careful source can do and a careless one cannot.
Certified rosters at the bottom. Digital agents on every floor above.
Field 3. Errors noted by the referring party
The Post found that the site credited Senator Jon Husted of Ohio with supporting a bill that passed before he took office. Copilot repeated it. The site merged two Robert Garcias, both running in California, into one profile. It listed wrong committee assignments. Sparse profiles stated that candidates had no campaign website when they did.
These are the Post's findings. The Desk has not reproduced them and has no view on the site's quality. It records the mechanism as the Post described it: the dossier contained a sequence in which a senator supported a bill, and the bill had passed before the senator arrived, and the chatbot read the dossier and agreed.
The senator arrived after the vote. The machine put him in the room.
Field 4. Volume
By the founder's own data, as reported by the Post, Microsoft's AI systems referred to the Ballot Brief more than 680,000 times in a 30-day span across September and October. The founder: "It's about providing as much information to the voter as possible but not being their sole resource."
The Desk has considered the word "sole." Seventy-five percent is not sole. The Desk confirms this arithmetically and has no further comment on the remaining quarter.
The Desk notes that the only count of the traffic in the file is the count the source kept. The instrument that generated the traffic has supplied no figure of its own.
Field 5. Governing remit
On 28 October 2026, two days before the Post's test, Politico reported that OpenAI had opened an Election Operations Center. It is led by Josh Lawson, OpenAI's head of civic and strategic risk, staffed across several teams, and will run until the votes are counted. Its work, as reported: monitoring real-time abuse of ChatGPT, handling reports of unexpected model behaviour, investigating suspected violations of the usage terms. Voters asking about registration or polling places are routed to Democracy Works; those asking about results are routed to the Associated Press. Prompts for photorealistic election images are, per reporting on the company's safeguards page, rendered as cartoons. Lawson told Politico that the aim on images is to "degrade the evidentiary value of an image."
As Politico described it, the centre concentrates on OpenAI's own products. It is not an arbiter of election information elsewhere online.
Everything the Post found was elsewhere online.
The Desk records that the remit and the finding share an edge. The remit covers the tool. The finding concerns what the tool read. The Desk has no proposal regarding the edge and observes only that the centre, as described, worked to specification, and that the specification stops where the citation begins.
BUREAU NOTE: A reader may picture a room, and the room is real: staff, screens, a line to the Associated Press. The Desk asks the reader to notice what the room is pointed at. It watches the instrument for abuse by users, for behaviour the makers did not expect, for breaches of the terms. It does not watch, because its remit does not include, the thing the instrument reads before it answers. The one place the newspaper looked is the one place the room was not built to look. The Desk considers the room well designed. It was designed to watch the chatbot, and the chatbot was not the problem. The chatbot was fine. The chatbot read a website.
Field 6. Prior
The ratio was on file before the election supplied the denominator.
In May 2026 a preprint, not peer reviewed, by Allaham and Diakopoulos (arXiv 2605.23684) put 712 real-world queries on politics, health and the environment to ChatGPT, Copilot, Gemini and Perplexity and classified the sources they cited. Approximately 16% of the 19,154 sources that could be scraped were classified as AI-generated, a figure the authors treat as a lower bound. Copilot's share was 27.8% of its cited sources, the highest of the four engines, and the highest in each of the three topics.
The May ranking is the October ranking. The Desk's own reading of the preprint is dated 2 November, which is also after the election supplied the denominator. The Desk files this against itself.
Field 7. Adjustments following publication
After the Post shared its findings, ChatGPT's citation rate for the site fell from 22% to 4%. Claude's rose. Microsoft said: "We've made additional updates to improve how Copilot responds to election-related questions," and said it had shifted Copilot to prioritise authoritative sources and official state election sites. Anthropic said that "ahead of the midterms, we've taken several steps to pressure test our models, strengthen our safeguards," and did not explain the movement toward the site.
Two adjustments, one in each direction. One statement from each party describing its safeguards as prior. The Desk records that the correction, where it occurred, was triggered by a newspaper sharing a spreadsheet with the companies, and notes what the newspaper is.
The newspaper is the thing the chatbot was replacing.
Field 8. Disposition
The third filter is sourcing. The news runs on a steady supply of credible sources, and whoever supplies them cheaply, reliably and in the right format becomes the news. That is the doctrine as written in 1988, and it has not aged. It has removed the humans.
The 2026 arrangement, as entered on this form: a fleet of agents builds a dossier from certified rosters and the open web. A chatbot, needing a source shaped like an answer, cites the dossier. A voter, needing an answer, asks the chatbot. The maker of the chatbot staffs a room to watch the chatbot. The newspaper, which the voter did not ask, reads the dossier and finds a senator in a room he had not yet entered.
Nobody chose the Ballot Brief. It runs on convenience, not conspiracy. There is no cabal in Redmond selecting a site from Miami Beach; there is a site that was there, in the right format, when a machine went looking for every race and what the candidates say. Provenance is not a crime. It is a field on a form, and the field was empty, and the chatbot does not ask.
The site built by agents was cited by agents and read by voters. The Desk has entered it.
BUREAU NOTE: The reader will notice that this form carries a field marked "compiled by," and that the Desk has filled it in once, in its own name, with its sources listed by date. The Desk does not regard this as a virtue. It regards it as the minimum, and records that for 3,100 counties, in the week before a national election, the minimum was supplied by the source and not by the instrument that cited it. The Desk closes the form. It does not close the file. The file is the mechanism, and the mechanism is open until somebody asks the chatbot where it got that.
