Politics

Fake Local News Sites Boost Democrats, Surface in ChatGPT

Fake Local News Sites Boost Democrats, Surface in ChatGPT
Table of Contents
  1. Deceptive Political Advertising via AI
  2. Network Structure and AI Optimization
  3. Chatbot and Search Engine Responses

A network of websites disguised as local news outlets is promoting Democratic candidates ahead of the midterms, with OpenAI’s ChatGPT frequently citing them when voters ask for election information, according to an NPR investigation. Digital security firm VisionHeight identified 37 websites in the network, including 10 pretending to cover local news that each promote a single Democratic candidate in competitive U.S. House races.

Deceptive Political Advertising via AI

One such site, the Kalamazoo Beacon, was cited at the very top of a ChatGPT response when NPR asked which House candidates in Michigan’s 4th Congressional District care about affordability, directing users to Democrat Sean McCann. While source code for the Kalamazoo Beacon describes it as reporting on southwest Michigan, NPR found it has published only 16 stories, all focusing on McCann’s policy positions and record.

Renée DiResta, an associate research professor at Georgetown University who studies online influence campaigns, told NPR that while political campaigns optimizing messages for AI is expected, posing as news outlets crosses into deception. DiResta noted that disclosing funding would be the ethical approach, adding that opposition research entities compete to ensure critical information is surfaced.

The operator behind the 37 websites remains unclear. NPR was unable to locate or reach the site operators, while impacted candidate JoAnna Mendoza of Arizona and the Democratic Congressional Campaign Committee denied involvement. Other promoted candidates and supporting Democratic super PACs did not respond to requests for comment.

Network Structure and AI Optimization

The 37 websites overlap in how they present candidates, featuring single-candidate news and opinion sites, as well as state and national roundup sites with names like Public Safety Democrats and Democratic Women in the House. VisionHeight and hybrid threat forensics firm co-founder Andreas Sjöstedt found that all 37 websites were established on September 15, with almost all registered within an eight-minute window. Hundreds of articles appeared in batches in September and October, indicating the operator may have utilized artificial intelligence to generate the content, according to the VisionHeight report.

The network of websites explicitly encourages AI scraping by allowing a long list of AI crawlers from major companies and formatting articles as questions and answers. Additionally, Alisha Ukani, a postdoctoral researcher at the University of Washington’s Center for an Informed Public, pointed out to NPR that every website contains an llms.txt file listing site articles in plain text, a feature mostly found on technology and business websites.

Chatbot and Search Engine Responses

To measure the group’s success, NPR tested 37 prompts across ChatGPT, Gemini, and Google search. ChatGPT cited at least one article from the network in 19 out of 37 responses, primarily drawing from the fake local news sites or state politics sites. For downballot races with limited credible reporting—a phenomenon known as a data void—ChatGPT cited the network’s Wolverine State Report every time when queried about five Michigan city council candidates.

Google’s products responded differently. Gemini did not cite any links from the websites and occasionally declined to answer, while Google search and its AI Overview excluded the network. Google spokesperson Davis Thompson told NPR that the company detected a policy violation when at least one site briefly ranked high in early October and took swift action.

OpenAI spokesperson Nate Evans stated that ChatGPT’s search tools cite a range of sources and that the company takes action to improve model responses when issues are highlighted. However, subsequent NPR queries showed the Kalamazoo Beacon was still cited as a source.

Patrick Warren, a professor at Clemson University’s Media Forensics Hub, told NPR that disguised political advertising has a long history through ad flyers resembling newspapers or ads mixed into search results, viewing AI chatbot citations without source evaluation as simply the latest evolution of the experience.

You May Also Like

Leave a Comment

Your email address will not be published. Required fields are marked *