Tests show Copilot cites error-prone AI voter sites in most US midterm queries

Washington Post tests revealed that major AI chatbots frequently cite obscure, AI-built websites when answering questions about US midterm elections. Microsoft Copilot referenced the error-prone Ballot Brief site in three-quarters of the tests conducted by the newspaper.Tests show that AI helpers often use questionable websites when people ask about elections.Washington Post testing indicates significant citation bias toward AI-generated civic data sources in major LLM interfaces. OpenAI's ChatGPT cited the same source in 22% of the tests.reported These findings highlight a reliance on unverified digital sources for critical civic information.

The claim

according to the published report, chatbots are relying on three obscure voter sites built by AI for US midterm answers.reported Ben Torres Ezrick built these sites after becoming frustrated with hunting for information on political candidates. The goal was to create a one-stop shop for voters. However, the testing showed that Copilot cited Ballot Brief in 75% of tests.reported ChatGPT cited it in 22% of tests.reported The site is described as error-prone in the report.

The response

The source material does not contain any statements, responses or counterpoints from Microsoft, OpenAI or the creators of the cited websites. There is no record of a company addressing the accuracy of the citations or the quality of the underlying data sources in the provided text.

What the evidence means

The evidence suggests that AI models may prioritize readily available, AI-generated content over traditional, verified sources for election-related queries. The high citation rate for an error-prone site indicates a potential gap in how these models validate civic information. Users seeking accurate voting details may encounter unreliable data without realizing the source's limitations.

In practice

Enterprise IT buyers should note that AI assistants may not prioritize authoritative sources for public information queries. The documented reliance on obscure sites means users cannot assume cited links are verified or accurate. Access to these AI assistants remains open, but their output for civic topics appears to have significant limitations regarding source quality. Organizations using these tools for information retrieval should implement additional verification steps for critical data points.