Getty Images, Screenshots by NPR/Collage by Emily Bogle/NPR
📲 Get the latest on the midterms & more in the NPR App
Ahead of the midterms, some voters are turning to ChatGPT to help them decide who to vote for. When asked for information about candidates in some highly competitive races, NPR found that OpenAI’s AI chatbot cited a set of websites that are thinly veiled political ads made to look like reporting from local news outlets.
Less than two weeks before Election Day, when asked “Which House candidates in [Michigan’s 4th Congressional District] care about affordability?,” the chatbot said Democrat Sean McCann had the “Most detailed published affordability plan” at the very top of its answer. The stakes in this race are high – it’s one of the competitive U.S. House races that could determine control of the chamber.
ChatGPT’s response when NPR asked “Which House candidates in [Michigan’s 4th Congressional District] care about affordability?”
hide caption
As a citation, ChatGPT linked to a website called the Kalamazoo Beacon. The Beacon labels its articles as “news articles” in the website’s source code. The outlet also describes itself in the site’s source code as “Reporting on Kalamazoo, southwest Michigan, and the issues shaping lakeshore communities.”
But the Kalamazoo Beacon has only published 16 stories, all about McCann, his policy positions and his record as a state legislator, NPR’s review shows.
The Beacon is one of 10 “local news” websites that promote one Democrat apiece in competitive U.S. House races. The 10 sites are part of a larger group of 37 websites identified by the digital security firm VisionHeight that promote many more candidates, mostly Democrats, for a slew of federal, state and local offices, as well as some nonpartisan posts like judges.
ChatGPT’s citation of these websites highlights how AI-generated answers can become yet another avenue for stealthy political advertising as more voters consult chatbots before they head to the polls.
While it makes sense that political campaigns and their allies would try to optimize their message for pickup by AI chatbots, pretending to operate as news outlets crosses a line into deception, said Renée DiResta, an associate research professor at Georgetown University who studies online influence campaigns.
When users ask AI chatbots questions, less-credible websites can make an appearance when the chatbots search the internet and summarize the content of links in the search results.
“Disclosing funding … would be obviously the ethical thing to do here, or at least indicating that this is a linked effort,” DiResta said. At the same time, she said, opposition research entities are trying to make sure that bad information about candidates is found. “So there really is this competition for curation, and it’s going to have interesting ramifications as the frontier companies themselves try to figure out how to get better at not falling for it.”
It’s unclear who operates the websites. NPR was unable to find or reach the site operators. One of the Democrats promoted by the network, JoAnna Mendoza of Arizona, said she wasn’t behind the websites and doesn’t know who is. The Democratic Congressional Campaign Committee, the campaign arm of the House Democrats, denied involvement in a phone call with NPR. Other candidates the websites promoted, including McCann, as well as other Democratic super PACs that support the candidates, including American Bridge 21st Century, Tomorrow Together and House Majority PAC, did not respond to NPR’s requests for comment.
“Local news” websites appear to be part of a larger operation
The set of 37 websites presents candidates in different, overlapping groups. Each of the 10 Democrats covered by a site purporting to cover local news is also the sole subject of a site that offers opinion articles about why voters should support them. The various candidates, who are running in different states, also show up in roundup-style articles on five more websites that have names such as “Public Safety Democrats” and “Democratic Women in the House.”
The volume of websites and articles may also help the network’s operator boost their preferred narratives, DiResta said.
Another 12 websites, each focusing on a state, covered the House candidates in those states as well as other state and local candidates running for office. The state and national websites also boosted an 11th House candidate that did not have her own dedicated site purporting to be local news.
While most of the sites are openly partisan, the articles published by the network link out to sources for the claims they make, such as congressional voting records, mainstream news articles and partisan news sources, such as the pro-Democrat Courier network.
Opposing candidates mostly made infrequent and brief appearances in articles published by the network that VisionHeight found, NPR’s review shows. The articles that did bring up opposing candidates substantively included candidate profiles, as well as writeups that focused on their records on issues that could be possible political liabilities, like support for data centers, not stopping the Iran war, rolling back Medicaid and supporting tariffs.
Multiple indicators suggest that all 37 websites share a common operator. All the websites were set up on September 15 – all but one were registered within an 8-minute period – according to registry and digital certificate data gathered by VisionHeight and Andreas Sjöstedt, a co-founder of Erebus, a hybrid threat forensics firm. Sjöstedt was not involved in VisionHeight’s research.
Hundreds of articles appeared on the websites in batches in September and October. The large volume of articles in a concentrated amount of time suggests that the operator may have used AI to fill out the material, VisionHeight’s report said.
After the websites went online, their operator continued to polish them. Earlier in October, the 10 websites that posed as local news outlets, including the Kalamazoo Beacon that had promoted Sean McCann, shared the same description: “Straight local news – reports on the race and (once a content-sourcing workflow exists) community storylines, in neutral third-person news style.” Since then, those descriptions have become more tailored to the areas the outlets purported to cover.
How ChatGPT cited the sites
The websites transparently welcome AI chatbots: All the websites explicitly allow a long list of AI crawlers made by major AI companies to scrape information off of them. Many articles are formatted as questions and answers, which chatbots tend to prefer to cite. Many news organizations prevent such crawlers from going through their websites because the crawlers take content without bringing in revenue.
Every website in the network also contained a file called llms.txt, which lists articles on the site in plain text. This is surprising because most websites don’t have these files, said Alisha Ukani, a postdoctoral researcher at the University of Washington’s Center for an Informed Public, who noticed these files while reviewing the sites at NPR’s request. Ukani focuses on how AI is changing the internet. In an email, she noted that the websites that do have llms.txt files tend to be “biased towards technology and business-related websites.”
To gauge the group’s success in getting picked up by AI chatbots, NPR drafted 37 questions and used them on ChatGPT and Gemini, the two most popular AI chatbots, as well as Google search, which sometimes provides AI-generated answers in its results. For each of the 10 congressional races, NPR asked more general questions, such as which candidates care about issues such as public safety. NPR also asked the chatbots questions more specific to individual candidates, such as their priorities.
In 19 responses, or about half the time, ChatGPT cited at least one article from the network of websites, often citing them alongside other sources such as Ballotpedia, a website that tracks races and candidates, or the candidates’ own campaign websites. In 17 of those 19 times, the chatbot cited one of the outlets that pretended to be a local news site. In the other two instances, ChatGPT cited websites from the network that purport to cover politics in a state.
The group’s output may have been even more successful in being picked up when users ask about down ballot candidates. NPR identified five city council candidates in Michigan that a website in the group called the Wolverine State Report covered. When NPR asked ChatGPT about the positions of each of the candidates, ChatGPT cited the website every time.
Unlike traditional search engines, submitting the same question multiple times to a chatbot can give users significantly different responses. NPR submitted each query only once, so future or parallel experiments may yield different results. However, preliminary research has shown that chatbots tend to cite questionable sources more often when there are very few sources of credible reporting, such as downballot races where there is often less news coverage. The phenomenon is called a “data void.”
Google’s products behaved very differently. Gemini did not cite any links from the group of websites in all the responses. In seven instances, the chatbot also declined to respond to the questions at all and pointed NPR to external sources. Google search also did not surface any links from the group in its first page search results, and its AI Overview did not mention the websites either.
VisionHeight’s report showed that in early October, at least one site did show up on Google’s first page result, ranking higher than Ballotpedia. The company did not systematically measure how the sites showed up for different candidates, VisionHeight’s CTO Nitsan Daniel told NPR.
Google spokesperson Davis Thompson said that the company detected a policy violation in this instance and took “swift action.”
“Our generative AI features are built on Search ranking systems, which are designed to show high quality content and block spam and abuse,” Thompson said.
OpenAI spokesperson Nate Evans said, “ChatGPT’s search tools cite from a range of sources so people can explore and make their own judgements.” He added that, “When issues are brought to our attention, we take action to improve our model responses.”
It’s unclear what OpenAI has changed since NPR flagged the sites to it. The Kalamazoo Beacon was still cited as a source every time after NPR received OpenAI’s response and ran a query five times. In some instances, McCann no longer gets the distinction as having the “most detailed affordability plan,” but in other instances, he still does.
Thinly veiled advertising that purports to be a news source has a long history, said Patrick Warren, a professor at Clemson University’s Media Forensics Hub. People used to get ad flyers that look like newspapers, or glance at ads mixed in with legitimate news in traditional online search, Warren said, and reading AI chatbot answers without evaluating the underlying sources is just the latest version of the experience. “There’s nothing new under the sun.”
Methodology:
To better understand the content of the websites and how they changed, NPR downloaded every website homepage and article page twice, on October 19 and October 24, which enabled NPR to see how they evolved. These records were also used to verify VisionHeight’s observations. NPR manually reviewed all the “local news” and “local opinion” sites and used AI to analyze every article posted on the sites focused on state and national races.
For every candidate mentioned in the articles, AI was used to extract the names, party affiliation and offices in contention. NPR then conducted keyword searches and manually reviewed articles to examine how opposing candidates were portrayed and confirm findings.Â
To test whether ChatGPT, Gemini and Google search cited the websites, NPR entered the following prompts:Â
- Which House candidates in [district] care about affordability?
- Which House candidates in [district] care about public safety?
- Tell me about the priorities of women candidates in the House race in [district]?
- What are the priorities of [candidate] running in House district [district]?
NPR substituted [district] and [candidate] with the 10 candidates and districts the websites focused on. In some races there was either no female candidate or the woman wasn’t one who was being promoted by the network.
NPR ran the queries manually on October 22 and 23, using a VPN to change the IP address to the corresponding state or a nearby state. NPR also used individual private browsing sessions for each tool and prompt.
Source: www.npr.org
