ChatGPT cites fake local news sites for 10 House races

ChatGPT cited at least one article from a network of 37 websites in 19 of 37 tested queries regarding U.S. House races, according to a review conducted by NPR. These websites are designed to mimic local news outlets but primarily promote Democratic candidates. The network includes 10 sites that specifically pose as local news, each covering a single congressional district.

In a race in Michigan's 4th District, the chatbot identified Democrat Sean McCann as having the most detailed published affordability plan. It linked to the Kalamazoo Beacon, a site that describes itself in its source code as reporting on southwest Michigan. The Kalamazoo Beacon has published only 16 stories, all of which focus on McCann and his record as a state legislator.

Digital security firm VisionHeight identified the larger group of 37 sites. These outlets promote candidates for federal, state, and local offices, as well as nonpartisan judicial posts. The sites were all registered on September 15, with all but one established within an eight-minute window. Hundreds of articles appeared on these sites in batches during September and October, suggesting the use of AI to generate content.

Renée DiResta, an associate research professor at Georgetown University who studies online influence campaigns, said that pretending to operate as a news outlet crosses a line into deception. She noted that disclosing funding or indicating a linked effort would be the ethical approach. DiResta explained that there is a competition for curation as AI companies try to improve their ability to filter sources.

The websites explicitly allow AI crawlers to scrape their content. Many articles are formatted as questions and answers, a structure that chatbots tend to prefer. Each site also contains a file called llms.txt, which lists articles in plain text for machine reading. Alisha Ukani, a postdoctoral researcher at the University of Washington, noted that such files are rare outside of technology and business sites.

NPR tested the influence of these sites by asking 37 questions of ChatGPT, Gemini, and Google search. ChatGPT cited the network in 19 responses. In 17 of those instances, it cited one of the 10 sites posing as local news. In the other two, it cited state-level political sites. When asked about five city council candidates in Michigan, ChatGPT cited the Wolverine State Report every time.

Google's products behaved differently. Gemini did not cite any links from the group in its responses. In seven instances, Gemini declined to answer and pointed users to external sources. Google search did not surface links from the group on its first page, and its AI Overview did not mention the websites. However, VisionHeight found that at least one site appeared on Google's first page in early October, ranking higher than Ballotpedia.

OpenAI spokesperson Nate Evans stated that the company's search tools cite a range of sources to allow people to make their own judgments. He added that the company takes action to improve model responses when issues are brought to its attention. It remains unclear what changes OpenAI has implemented since NPR flagged the sites, as the Kalamazoo Beacon was still cited in repeated queries after the company's response.