37 fake news sites boost Democratic candidates in AI answers

A network of 37 websites posing as local news outlets influenced artificial intelligence chatbots to promote specific Democratic candidates in competitive House races. Less than two weeks before the election, these sites cited in AI responses helped shape voter information on key congressional contests.

NPR testing found that ChatGPT frequently cited these sites when users asked about candidates in 37 races. The sites are part of a larger operation identified by digital security firm VisionHeight. The network promotes mostly Democrats for federal, state, and local offices, including nonpartisan positions like judges.

The Kalamazoo Beacon serves as a prime example of the operation. The site labels its content as news articles in its source code. It describes itself as covering Kalamazoo, southwest Michigan. However, the site has published only 16 stories. Every article focuses on Democrat Sean McCann, a candidate for Michigan's 4th Congressional District.

VisionHeight and forensics firm Erebus determined that all 37 sites were created on September 15. All but one were registered within an eight-minute window. Hundreds of articles appeared in batches during September and October. This volume suggests the operator used AI to generate the material. The sites allow AI crawlers to scrape their content, a practice many traditional news organizations restrict.

ChatGPT cited at least one article from this network in 19 of 37 test queries. In 17 of those instances, the chatbot referenced one of the sites posing as local news. Google's Gemini did not cite any links from the group in its responses. Google search also failed to surface the sites on its first page of results, although one site briefly ranked higher than Ballotpedia in early October.

Google spokesperson Davis Thompson said the company detected a policy violation and took swift action. He stated that generative AI features rely on search ranking systems designed to block spam and abuse. OpenAI spokesperson Nate Evans said the company takes action when issues are brought to its attention to improve model responses.

Renée DiResta, an associate research professor at Georgetown University, noted that pretending to operate as news outlets crosses a line into deception. She said disclosing funding or indicating a linked effort would be the ethical standard. The sites often link to credible sources for their claims, such as voting records and mainstream news articles, to bolster their appearance.

The operation remains anonymous. NPR could not identify or reach the site operators. Democrat JoAnna Mendoza, one of the candidates promoted by the network, said she was not behind the websites and did not know who was. The Democratic Congressional Campaign Committee denied involvement in a phone call. Other candidates and supporting super PACs did not respond to requests for comment.

Alisha Ukani, a postdoctoral researcher at the University of Washington, noted that the sites contain a file called llms.txt. This file lists articles in plain text to help AI models read the content. Most websites do not have these files. Ukani observed that sites with llms.txt files tend to focus on technology and business topics, making this political network unusual.