Copilot Cited Ballot Brief in 75% of Tested Midterm Replies

TikTok named the organizations supplying its 2020 election guide. In tests of U.S. midterm questions, Microsoft’s Copilot cited the obscure, error-prone AI-built site Ballot Brief in 75% of responses, without a disclosed partnership appearing in the available record. Both systems put voting information in front of users. Only one announced its source relationship before the user asked.

Answer engines are turning citations into a new layer of civic authority: by choosing which pages sit beneath an answer, platforms can give a publisher practical reach before its provenance, correction process, or accountable institution is legible to voters. Publishers consequently compete not only to publish election information but to become the sources an answer engine selects, while platforms inherit the work of deciding which sources deserve that position.

TikTok named its suppliers before it offered election answers

Digital election guides were never free of platform judgment. In 2020, TikTok partnered with BallotReady, SignVote, and other organizations to provide candidate and voting information in its U.S. elections guide. TikTok still chose which organizations belonged in the app, but it made that choice visible as a partnership. A voter, journalist, or election official could identify an information provider and ask what it had supplied.

Earlier civic-information providers also had organizational histories that could be examined independently of any one answer. A 2016 profile identified Debra Cleaver as the founder of Vote.org, described the nonprofit’s Y Combinator backing, and explained its use of text messages to increase turnout. That history did not guarantee that every message was correct. It gave scrutiny an address: an organization, a method, and people responsible for the work.

Ballot Brief’s appearance in the chatbot tests establishes a different kind of relationship. The supplied record shows citations, not a disclosed arrangement between Ballot Brief and either chatbot. Nor does the absence of a recorded partnership prove that none exists. What matters to the voter receiving an answer is that the platform can select a source without first presenting the kind of named institutional relationship TikTok presented at the entrance to its guide.

Copilot’s earlier election errors made retrieval an operating question

Researchers reported in December 2023 that Copilot answered some political questions with conspiracies, misinformation, and outdated or incorrect information. That warning preceded the Ballot Brief tests by nearly three years. It identified a problem broader than a chatbot’s ability to write a fluent sentence: an election answer also depends on what information the product obtains, which claims it carries forward, and whether it knows when not to answer.

In the 2026 midterm-query tests, Copilot cited Ballot Brief in 75% of responses, while ChatGPT cited it in 22%. The Washington Post described the site as error-prone and found chatbots drawing on three obscure AI-built voter-information sites. Copilot gave Ballot Brief markedly more exposure than ChatGPT did on the tested questions. Neither test establishes how widely the site appears in answers elsewhere.

Ballot Brief appears in one article in the available coverage from 2024 through 2026. That is a measure of the available coverage, not an audit of the site’s full history or evidence that voters changed their decisions. The tests establish how often a little-documented source appeared in particular answers. They do not establish how many voters saw those answers, clicked the citations, or acted on them.

Cheap publishing gives retrievers more pages to judge

The supply of plausible source material changed before the latest election tests. In 2023, the Washington Post documented AI-generated material spreading across the web, including recipes, reviews, and books. In March 2026, WordPress.com said AI agents could draft, edit, and publish on customer sites, as well as update metadata and manage comments. Neither report establishes that those tools produced Ballot Brief. Together they show why a current-looking page is becoming cheaper to create and maintain.

An answer engine must make several decisions after a user submits a question: which pages to retrieve, which passages to use, how to reconcile conflicting claims, and what to cite. A polished page can help at the first step without resolving the later ones. A date in a page template is not necessarily the date an election instruction was checked; two pages repeating a claim are not necessarily two independent confirmations. The useful test is whether the cited organization can account for the particular claim, jurisdiction, and revision the answer relies on.

Companies have made retrieval an operational product layer. Contextual AI, for example, has marketed end-to-end retrieval-augmented generation for enterprise answers alongside an emphasis on governance and privacy controls. An enterprise can define a document collection and decide who may change it. An open-web election answer faces a harder version of the same source problem: the publisher and the answer engine may be separate institutions, and the voter sees the assembled response rather than their working relationship.

Publishers can optimize for selection before voters can inspect it

Generative Engine Optimization firms including Profound and Bluefish AI sell businesses tools to improve how their content appears in AI search summaries. That is a rational response to a changed distribution surface. Brave Search places synthesized answers above organic results for informational questions; Google has expanded AI Mode and AI Overviews to include firsthand perspectives from social media and forums, presented as “Expert Advice.” Different products choose different source pools and different positions for the resulting answer.

Google’s inclusion of forum perspectives can add useful firsthand knowledge. A forum post, though, earns its place by answering a question, not by passing the same institutional checks as a named election-information partner. Likewise, a publisher can make a page easier for a system to retrieve without establishing who verifies its voting instructions. The commercial techniques for reaching the answer box and the civic tests for belonging there are separate tasks.

Publishers already face a related reversal when AI summaries deliver information without a visit to the originating site. The AI search publisher-licensing bargain concerns payment and traffic when an answer replaces a click. Election information adds a different cost to the same interface choice: a user may act on the answer before visiting any source at all.

Google’s accuracy record argues against judging the system by its worst screenshot

Google says the vast majority of AI Overviews provide high-quality information and that many viral failure examples involved uncommon or doctored queries. An analysis of Gemini 3-based AI Overviews estimated roughly 90% accuracy. Those are substantial counterweights to a story told entirely through spectacular errors. A 2024 review also found AI-search citations helpful for exploratory questions, even as it rated the tools worse than Google for navigation and mixed for informational searches.

None of those findings measures the reliability of an answer about a particular voter’s election question. A broad accuracy estimate cannot establish performance on local, time-sensitive civic queries, just as the Ballot Brief tests cannot establish that every chatbot makes the same source choice. Google’s wider source pool and the Copilot–ChatGPT gap both make product-specific, query-specific testing more useful than treating a citation—or a failure screenshot—as a verdict on the whole category.

Election answers need a correction path, not a longer footnote

Providers have choices besides answering directly. OpenAI reported that ChatGPT directed roughly two million election-news requests elsewhere on November 5 and 6 during the 2024 U.S. election, and rejected more than 250,000 election-related image requests. Ahead of the 2026 midterms, OpenAI also announced an election “war room” to identify misinformation and disruption efforts. Referrals, restrictions, and active monitoring show that the product operator can change the path a civic question takes.

For an answer it does provide, the operator needs a way to inspect more than the presence of a citation: who published the source, which jurisdiction and date its claim covers, whether the relevant text has been corrected, and who is responsible when the answer conflicts with an authoritative record. A human reviewer matters not merely as a last defense against an error, but as someone accountable for the decision to answer, refer, or withdraw an answer. Testing those decisions against the questions voters actually ask would reveal failures that a general accuracy score cannot locate.

TikTok put BallotReady’s name at the entrance to its election guide. Copilot put Ballot Brief’s name beneath an answer. When that answer needs correcting, the voter needs to find more than the link.