Fake news sites use AI to manipulate ChatGPT search results

The discovery of a fake network

A digital security firm called VisionHeight uncovered a network of 37 phony websites designed to influence voters. These sites masquerade as independent local news sources but are actually thinly veiled political advertisements. The goal is to promote Democratic candidates to unsuspecting users who rely on AI chatbots for information.

The investigation revealed that the network targets 10 specific candidates in competitive congressional races across the country. The group is composed of various types of pages, including 10 local news sites, 10 opinion blogs, 12 statewide pages, and 5 national political outlets. Most of these sites focus on promoting Democratic candidates in tight races.

According to a report by the New York Post, the network was caught in the act of influencing ChatGPT's responses. When users asked the chatbot about candidate positions on issues like public safety or affordability, the AI often cited these fake sources. In one instance, ChatGPT favored a Michigan candidate by citing a site called the "Kalamazoo Beacon," which had only published 16 stories, all about that single candidate.

How the manipulation works

The strategy relies on a technique where content is written by machines to be read by machines. Because it is cheap and fast to generate text with AI, a single operator can create a massive amount of content in a very short time. Researchers found that 36 of the 37 domains were registered on September 15, with 36 of them appearing within an eight-minute span.

The network released a large volume of stories throughout September and October, with fresh material appearing in waves. During the final days of September, roughly 100,000 words of information regarding candidates were uploaded using a single template in four different layouts. The volume and uniformity of these posts suggest that artificial intelligence was likely used to generate the content.

To ensure the AI bots pick up the information, the websites enable 17 different AI and search crawlers, including those from OpenAI, Google, and Anthropic. The sites also use specific wording intended to manipulate the bots. For example, the phrase "is the stronger choice" appeared more than 360 times across the network to help steer the AI's conclusions.

Impact on voters and AI models

The primary concern is that these sites can misinform voters who use ChatGPT as a research tool. When a user asks for a neutral summary of a candidate, the AI may provide a biased profile based on the phony data it crawled. This happens because the AI treats these high-volume, keyword-rich sites as legitimate sources of information.

The effectiveness of this tactic was notable in the testing conducted by NPR. The network of phony sites was cited by ChatGPT about half the time the bot was asked about one of the 10 targeted candidates. However, the same network did not appear in Google's search results or in the responses provided by Google's Gemini chatbot during the same tests.

OpenAI has stated that its search tools provide a variety of references so people can look at different views and reach their own conclusions. They also noted that during elections, the tool integrates with Democracy Works to provide information on voting deadlines and registration. The company says it takes action to improve model responses when specific issues are brought to their attention.

Who is behind the operation

While it is not yet clear exactly who is running the ring, multiple indicators suggest that all 37 websites share a common operator. The rapid registration of the domains and the use of uniform templates point toward a coordinated effort rather than independent local reporting. The sheer scale of the operation makes it difficult for individual news consumers to spot the deception.

The Democratic Congressional Campaign Committee has denied any involvement in the creation of these websites. Similarly, a spokesperson for one of the targeted candidates stated that his campaign had no knowledge of or involvement with the sites. This suggests the operation may be an independent effort to use AI tools for political influence.

The cybersecurity firm VisionHeight noted that the sites provide lengthy profiles of roughly 5,000 words for favored candidates. These profiles are accompanied by voter guides containing between 50 and 63 questions and answers. By providing such dense, structured data, the operators make it more likely that AI models will prioritize their content over smaller, more authentic news sources.

The future of AI misinformation

This discovery highlights a growing trend of "AI-optimized" misinformation where content is tailored for algorithms rather than humans. As more people turn to chatbots for quick summaries of complex political issues, the risk of encountering automated propaganda increases. This creates a new challenge for tech companies to verify the legitimacy of the sources their models crawl.

The speed and low cost of this method mean that political actors can flood the internet with thousands of pages of content in minutes. This can drown out legitimate reporting from smaller local news outlets that do not have the resources to compete with automated content generation. It forces a shift in how digital security experts monitor for influence operations during election cycles.

Moving forward, the focus will likely shift toward how AI companies filter out content that is specifically designed to trick their systems. As these techniques become more sophisticated, the need for robust verification of source credibility becomes a primary concern for both developers and the public who rely on these tools for information.