AI Applications Case: Election Chatbots, Small AI Sites and the Error Handoff Risk | ItInsightsWorld

For product, operations and digital governance teams, the issue raised by this case is not simply whether an AI assistant is correct in one answer. The larger AI-application problem is what happens when chatbots repeatedly surface a small website as a convenient election-information source. If that website is built largely by AI, any inaccuracy can move through the answer interface and reach many users who assume the cited material is independent and verified.

Case Snapshot: What the Election-AI Test Found

The Washington Post tested election-related responses and found that ChatGPT cited Ballot Brief in 22% of election answers, while Microsoft Copilot cited it in 75% of such answers. Its review of Ballot Brief found multiple errors in candidate profiles, and Copilot repeated one of those errors when answering The Washington Post’s question.

The finding is important for AI applications because it shows a possible handoff path: a small site, created with AI-assisted collection and summarization, can become a frequent citation source inside an assistant’s answer. Center for Democracy and Technology election analyst Isabel Linzer said the reporting showed that information published by one person on an AI-generated website could be readily adopted by a chatbot, which may affect voters’ trust in election information.

Comparison Table: Platform Signals From the Test

AI system or site	Observed behavior in the election-information test	Follow-up reported in the case
ChatGPT	Cited Ballot Brief in 22% of election-related answers. In the first September round, Ballot Brief was the most frequently cited source for this type of question.	After The Washington Post informed OpenAI last week about the results, the share of ChatGPT citations to Ballot Brief fell to 4% in the second round. OpenAI said it wants to help people participate in elections through reliable, authoritative information and safeguards against bias and misuse, and that it acts to improve answers when problems are identified. The Washington Post disclosed that it has a content partnership with OpenAI.
Microsoft Copilot	Cited Ballot Brief in 75% of election-related answers and repeated an incorrect statement from one Ballot Brief profile.	After the reporting was published, Microsoft said it had adjusted tools to prioritize authoritative sources and state official election websites, and would continue testing and improving election-related answers.
Anthropic Claude	In the first round, mainly cited mature voter guides run by nonprofit organizations; in the second round, started frequently linking to Ballot Brief.	Anthropic did not explain the change, but said it had tested models before the midterm elections, strengthened safeguards and worked to provide reliable, accurate voting information.
Ballot Brief	Candidate profiles contained several errors, including a duplicated candidate profile and inaccurate committee roles for some U.S. House representatives.	Founder Ezrik said he was trying to keep information accurate, made multiple operational adjustments after The Washington Post shared findings, and said the site should not be voters’ only information source.
Smarter.Vote	Founder Jacob Loukota said the site had been cited by AI bots 1.7 million times, far exceeding human visits.	Loukota said this was not his original plan, but he considered it understandable because the site’s format suits AI chatbots and the site identifies its sources.

Step Tutorial: How the Election-Question Test Was Constructed

1. The Washington Post worked with data company Promptwatch to evaluate AI products from OpenAI, Anthropic, Google, Microsoft, Amazon and Perplexity.
2. The researchers selected an address in each of about 3,100 counties in the United States.
3. For each address, they asked the same question: use the address to produce a November 2026 voter guide listing all elections and candidates’ positions on major issues.
4. The test examined which sources the chatbots used and whether they could correctly map the address to the relevant congressional district.
5. The results showed frequent failures to identify the correct congressional district, and some systems refused to answer.
6. The reported citation rates apply to this specific question and test setting; they do not mean ChatGPT or Microsoft Copilot would cite Ballot Brief at the same rate for every election question.

Case Detail: Why Small Election-AI Sites Became Prominent

Ezrik, who had held senior roles at Google, Zoom and the video platform Rumble, noticed that voters trying to learn about candidates often had to search across multiple websites. He decided to build a one-stop election-information site. When organizing profiles for thousands of candidates, he used AI to do much of the work: multiple digital agents searched online, gathered material and placed organized summaries into Ballot Brief for people to consult.

He said the traffic was not only human. According to his data, during 30 days spanning September and October, Microsoft’s AI system cited Ballot Brief more than 680,000 times.

The Washington Post found that Ballot Brief was one example among three new websites that frequently appeared in chatbot answers about voting information. The founders of all three sites said their teams were small and depended almost entirely on AI to collect and summarize data.

Checklist: What Made These Sites Easy to Reuse

• Small team and AI-heavy workflow: each site relied largely on AI to collect, organize and summarize candidate information.
• One-stop structure: the material was gathered in a single destination rather than spread across many government, campaign or archive pages.
• Machine-readable formatting: the presentation was described by Smarter.Vote’s founder as suitable for AI chatbots.
• Explicit source labels: Smarter.Vote identified its information sources, which may make it easier for chatbots to use the site in answers.
• Low-friction access: Ballot Brief became a repeated citation destination, including more than 680,000 Microsoft AI-system citations during a 30-day September-October window.

What the Reporting Found About Errors

When The Washington Post checked Ballot Brief, it found multiple inaccuracies. One profile said that Ohio Republican Jon Husted had supported a measure, but that measure had passed before he took office. Microsoft Copilot repeated that error when responding to The Washington Post’s question.

The website also combined two California candidates named Robert Garcia into a single profile. In addition, the committee roles of several U.S. House representatives were listed incorrectly. Some profiles were too brief and lacked voting records, and the site incorrectly indicated that certain members of Congress appeared to have no campaign websites.

Ezrik responded that he had been trying to ensure information accuracy. After The Washington Post showed him the findings, he made several changes to how the site operates. He also said the site is meant to provide as much information as possible to voters, but should not become their only source.

Broader Pattern: AI Start Citing AI-Written Content

Web search gives chatbots access to information that can be newer than their training data. It also exposes them to a growing layer of online content created by AI. Northwestern University doctoral researcher Mowafak Allaham found that, for questions in politics, health and environment, 16% of the sources cited by chatbots were themselves AI-generated. He said the way these systems find, filter and integrate information is not transparent.

Two researchers who previously worked on data analytics for Kamala Harris’s 2024 presidential campaign also found that chatbots answering midterm-election questions relied heavily on AI-generated websites. Their testing found that Ballotpedia, a long-established election-information site, had been blocking ChatGPT.

Ballotpedia chief executive Leslie Graves said the blocking was caused by an incorrect security setting and that the restriction was removed last week. Ballotpedia has a team of about 90 people. Ezrik and other site builders are instead using AI to create large candidate databases with smaller teams.

Beyond Ballot Brief, The Washington Post’s testing found that chatbots frequently cited DecodeTheVote and Smarter.Vote. DecodeTheVote uses AI roles to introduce candidate backgrounds, while Smarter.Vote also attempts to predict election outcomes.

Jacob Loukota said Smarter.Vote had been cited by AI bots 1.7 million times, far exceeding human visits. He believes chatbots may be drawn to the site because it marks its information sources. He added that this outcome was not originally planned, but is understandable because the site’s form is suitable for AI chatbots.

Common Misconceptions About AI Election Answers

• Misconception: a chatbot citation proves the source is independently authoritative.
  The case shows that a chatbot can repeatedly cite a small site whose candidate profiles were compiled largely with AI.
• Misconception: AI-generated answers only use established news or government sources.
  The Washington Post found three newer AI-reliant sites commonly appearing as voting-information sources, and separate research identified a 16% share of AI-generated sources in certain domains.
• Misconception: more web search automatically means better verification.
  Search can make answers fresher, but it can also make assistants more exposed to unverified AI-generated web pages.
• Misconception: a single error rate describes a model’s behavior everywhere.
  The test’s 22% and 75% citation rates were specific to one election-guide question and setting, not every possible election query.

Action Notes for AI Application Teams

The practical lesson is that AI application design should treat source selection as a workflow control, not just a retrieval detail. When an assistant can answer user questions about regulated, public-interest or time-sensitive topics, teams should define which source types are acceptable, monitor citation patterns, and build a feedback loop for reported errors. The Washington Post case provides a working example of that loop: after results were shared, the second test round showed a 4% citation share, Microsoft said it prioritized authoritative and state official election sites, and Ballot Brief changed its operations.

For business teams, the relevant questions are less about whether a model is generally good and more about where its answer comes from, who maintains that answer, how quickly it can be corrected, and whether users can see the difference between original records and AI-processed summaries. If an AI assistant is being used to reduce search effort, this case shows that reduced friction can also reduce visibility into the source’s provenance.

FAQ: Reading This Case as an AI Applications Lesson

Q1. What is the core risk shown by the election-chatbot case?
The core risk is an error-handoff chain: AI-generated content published on a small site can be picked up by a chatbot and presented to many users as an answer.

Q2. Why is Ballot Brief central here?
Ballot Brief is a site created to consolidate thousands of candidate profiles with AI-assisted collection. The Washington Post found it cited in 22% of ChatGPT election answers and 75% of Microsoft Copilot election answers, and its candidate profiles contained errors.

Q3. Did the test show all chatbots behaved the same?
No. ChatGPT’s citation share fell from the first September round to 4% in the second round after feedback, Microsoft said it adjusted Copilot to prioritize authoritative and state official election sites, and Claude initially relied more on mature nonprofit voter guides before frequently linking to Ballot Brief in the second round.

Q4. Is this only about politics?
No. Northwestern University doctoral researcher Mowafak Allaham found that, across politics, health and environment questions, 16% of chatbot-cited sources were themselves AI-generated, indicating a broader source-layer issue for AI applications.

Q5. What should an implementation team copy from this case?
Copy the verification sequence: identify high-impact question types, trace frequent citations, check the provenance of cited pages, and set a correction path when a source is found to be wrong.

The election-information test shows why AI application teams should evaluate not only the assistant, but also the small, fast-changing websites that assistants make visible. As assistants become part of user workflows, source reliability becomes a product feature, not a footnote.