The claims, checked
On Nov. 1, an X post by Jonathan Frye summed up an NPR investigation and asked who paid for what it found. A reply under the post argued that AI chatbots lean toward one party by design. Profoundd checked each factual claim against NPR's report, the security firm's original findings, registry and campaign finance records, and the 37 sites themselves. The last line of Frye's post is cut off in the copy Profoundd saw.
- "37 political websites. Registered together. Built to look like news." Confirmed. Registry records show all 37 addresses registered on Sept. 15 within 22 minutes, one race at a time in a fixed order: one builder working from one list, though that does not show who paid. Twenty-two present themselves as news outlets: ten hometown papers, which described themselves in their code as local news in "neutral third-person news style," and twelve statewide sites. The other 15 are openly partisan opinion and issue sites. None names an owner.
- "The websites promoted mostly Democratic candidates." Confirmed. Each of the ten look-alike local papers promotes one Democratic House candidate; the other sites add an 11th, and their down-ballot picks match a political committee's voter guide (see below). Some pages profile Republicans in a neutral tone, among them the Republican nominees for secretary of state in Arizona and Michigan, but Profoundd found no page promoting one in the 37 sites' article lists.
- "many were designed to make their content easy for AI chatbots to find." Confirmed; all 37 were. Every site carries the same file inviting 17 named AI and search crawlers in, and a plain-text list of its articles for AI tools. The Kalamazoo Beacon publishes six of its ten articles a second time as "Frequently Asked Questions."
- "In NPR's testing, ChatGPT cited the network in 19 of 37 answers." Confirmed, as NPR's result. NPR asked each of its 37 questions once, on Oct. 22 and 23, and notes that repeated tests could differ.
- "Some of those supposed local news sites barely covered anything except one candidate." Confirmed. The Kalamazoo Beacon's own article list has 16 entries, every one about Democrat Sean McCann, who is running in Michigan's 4th District.
- Trump "signed an executive order demanding truthfulness and ideological neutrality from AI purchased by the federal government," and "that order doesn't directly solve this problem." Confirmed. The July 2025 order sets "truth-seeking" and "ideological neutrality" as conditions for the AI models federal agencies buy, which can include ChatGPT under a 2025 government-wide deal. It does not reach the consumer chatbots voters use, or who runs political websites.
- "We don't have proof that Democratic Party leaders directed it. We don't even know the operator yet." Confirmed. No one has publicly identified the operator. The House Democrats' campaign committee, Arizona candidate JoAnna Mendoza and McCann's campaign have said they were not involved.
- "Who paid for all 37 websites, and who gave the orders?" Unanswered, with two leads. The sites carry no "paid for by" notice, and the registrar's record hides the owner, showing only a registrant in California. The statewide sites' down-ballot candidate lists come from a voter guide paid for by the State + Local Election Alliance, a federal political committee, which shows where the lists came from, not who built the sites.
- The reply: "This is not new," because AI models are "trained on legacy media and other biased sources," their guardrails point to a preference "By design," and "Its already built into the LLM's to lie for one party." The core holds; the design to lie is not shown. Independent studies since 2023 find that most chat models lean left of center, trace part of the lean to how models are tuned with human ratings, and find that users of both parties see it. None shows that anyone designed the lean or built models to lie.
What NPR and the security firm found
Less than two weeks before Election Day, NPR asked ChatGPT which House candidates in Michigan's 4th District cared about affordability. The chatbot put Democrat Sean McCann first, crediting him with the "Most detailed published affordability plan," and cited the Kalamazoo Beacon, a site that calls itself local news but has published only articles about McCann. NPR's Oct. 30 investigation, by Huo Jingnan, traced the Beacon to a network of 37 sites first identified by VisionHeight, a company that sells threat-detection software. Who runs the network is unknown. NPR could not find or reach the operator. The Democratic Congressional Campaign Committee denied involvement, and JoAnna Mendoza, an Arizona Democrat the sites promote, said she was not behind them and does not know who is. McCann did not respond to NPR, but his campaign told the New York Post it "has no knowledge of or involvement with these websites." Other promoted candidates and three super PACs that back them, American Bridge 21st Century, Tomorrow Together and House Majority PAC, did not respond to NPR.
Ten sites pose as hometown newspapers, each devoted to one Democratic House candidate, and ten opinion blogs urge a vote for the same ten. Twelve statewide sites cover those races alongside state and local contests, and five national sites carry names such as "Public Safety Democrats" and "Republican Incumbent Watch." VisionHeight counted the phrase "is the stronger choice" more than 360 times across the network and about 100,000 words of candidate material posted on a single day, Sept. 30, and concluded the sites were most likely written by AI. NPR, which downloaded every page twice in October, found most of the sites openly partisan. Their articles link to sources such as voting records, mainstream news and the pro-Democratic Courier network, and opponents appear rarely, mostly in pieces on their possible liabilities.
The sites were built to be read by machines. Each invites 17 named AI and search crawlers, including those of OpenAI, Google, Perplexity and Anthropic, whose Claude model Profoundd uses to research and draft its stories, this one included. Each also carries an llms.txt file, a plain-text list of its articles for AI tools, which most websites do not have, a University of Washington researcher told NPR. In NPR's test, ChatGPT cited the network in 19 of 37 answers, often alongside sources such as Ballotpedia or the candidates' own sites, and 17 of those times through one of the look-alike local papers. It cited the network's Michigan site every time NPR asked about five city council candidates. Google's Gemini cited none of the sites and declined to answer seven of the questions, and Google search showed none on its first page. Google said it had detected a policy violation and taken "swift action," without saying when; VisionHeight had found a network page ranked on Google's first page, above Ballotpedia, in early October. OpenAI said ChatGPT's search cites "a range of sources so people can explore and make their own judgements" and that the company acts on problems brought to its attention. When NPR ran a query five more times after OpenAI's reply, the Beacon was cited every time.
What Profoundd checked for itself
On Nov. 2, Profoundd pulled the registration record for each of the 37 addresses from Verisign, which runs the .com registry, and from the registrar, and fetched each site. All 37 were registered on Sept. 15 between 5:40 and 6:02 p.m. Eastern time. After the first, the rest followed in under eight minutes, most of them 11 to 17 seconds apart, in a fixed order: each House race's news site and opinion blog together, then the statewide sites, then the national ones. All went through Cloudflare's domain registrar with the same pair of Cloudflare name servers, which is consistent with a single account, and all 37 are hosted by Netlify. Their robots.txt files, which tell crawlers what they may read, are identical apart from each site's address. All were still online on Nov. 2.
No site names an owner, an editor or a reporter; bylines read "Staff Report." There is no About page and no "paid for by" notice, only a copyright line and a corrections form with no email address. The home pages load no analytics or advertising code, which removes one common way researchers link anonymous sites to their owners. The registrar's public record hides the registrant's name and street address on all 37 domains. The one field it shows, the same on every one, is a registrant in California; the account holder supplies that entry and nobody verifies it, so it is a lead, not an identification.
The ten look-alike local papers promote Sean McCann (Michigan's 4th District), Paige Cognetti (Pennsylvania's 8th), Janelle Stelson (Pennsylvania's 10th), Bob Brooks (Pennsylvania's 7th), Christina Bohannan (Iowa's 1st), Sarah Trone Garriott (Iowa's 3rd), Rebecca Cooke (Wisconsin's 3rd), JoAnna Mendoza (Arizona's 6th), Amish Shah (Arizona's 1st) and Shannon Taylor (Virginia's 1st). The 11th House candidate, covered on the statewide and national sites, is Eileen Laubacher, the Democrat challenging Republican Rep. Lauren Boebert in Colorado. The statewide sites also back Democrats such as Keisha Lance Bottoms in Georgia and Nevada's secretary of state, Cisco Aguilar, carry neutral-toned profiles of some Republicans, and add general guides such as Texas state-park reservations. One Beacon page separates McCann from a former Illinois state senator of a similar name who faced a federal criminal case, men the page says are "sometimes confused."
A lead: where the candidate lists came from
The statewide sites' picks for local and judicial offices did not start with the network. Its Mississippi site says three judicial candidates "have a Fair Judge designation" and cites localvoterguide.com for each, and its Arizona judicial pages link to the same guide. The guide's pages say it is "Paid for by State + Local Election Alliance," a Washington-based committee registered with the Federal Election Commission as a PAC with a separate account for unlimited independent spending. In North Carolina, the guide marks 21 local and judicial candidates as its picks, and the network's statewide site covers 17 of them and none of the guide's 22 other candidates, by Profoundd's count. The alliance's website features quotes from Democrats, including former Wisconsin Democratic Party chair Ben Wikler and Nevada's Aguilar, one of the network's candidates, and its FEC reports show it raises money through ActBlue, the Democratic fundraising platform.
That shows where the lists came from, not who built the sites. The guide is public and names its sponsor, so anyone could have copied it. The alliance's own websites use a different registrar, name servers and host from the network's. Its FEC reports through Sept. 30, which cover the evening the sites were registered, show no itemized payments to Cloudflare or Netlify; a vendor paying on its behalf, or a payment after Sept. 30, would not appear there. And the alliance's own campaigning does not line up neatly with the network: in Pitt County, North Carolina, Greenville's Daily Reflector reported Oct. 28 that the alliance mailed a flyer for Democrat Blake Price, a candidate the network does not cover. Profoundd has not yet put the question to the alliance.
Who paid?
That is Frye's central question, and the public record does not answer it yet. Cloudflare and Netlify know whose accounts these are and could be asked, or compelled through legal process, to say. Federal Election Commission rules require a "paid for by" notice on a political committee's public websites and on ads paid to appear on other people's sites, but the FEC's definition of a regulated "public communication" leaves out most other internet content, so whether these sites needed a notice depends on who is behind them. If a federal committee paid for the work, its FEC reports would have to show the spending, usually under general labels such as consulting or digital services. Records like these let Columbia University's Tow Center trace over $14 million spent in 2021 and 2022 on Metric Media, a Republican-aligned network of local-news look-alikes. Beyond the alliance, Profoundd has not yet searched the 11 campaigns' and the super PACs' reports for payments that could match.
Is the tactic new?
Partisan sites posing as local news are years old and used by allies of both parties. In 2022 NewsGuard, a company that rates news sites, counted 1,079 sites in the largest network, the Republican-aligned Metric Media, and tracked Democratic-aligned ones such as the American Independent and Courier Newsroom. Chatbots cited such sites before this network existed. In an audit by two NewsGuard analysts published by Politico on Sept. 25, seven chatbots answering questions about toss-up races cited a sample of 12 partisan look-alike sites, six from each side, in 48% of 168 answers, and only one answer told the user the source was partisan. ChatGPT did so most often, in 71% of its answers. What VisionHeight says it has not seen before is an American election network built from the start for AI tools rather than human readers. A Clemson University researcher gave NPR the long view: "There's nothing new under the sun."
Do AI models lean toward one party?
The reply argues this is old news because AI models are "trained on legacy media and other biased sources" and their guardrails, the rules added in training to steer and limit answers, push them toward one party "By design." Independent research supports the core of that. Training sets draw heavily on news sites; in a 2023 Washington Post analysis of one widely used set, The New York Times was the fourth-largest source. A 2023 University of Washington and Carnegie Mellon study found that models absorb the political leanings of their training text and carry them into tasks such as flagging hate speech. On the guardrails, a 2023 Stanford study found that models tuned with human ratings leaned left, and a 2024 MIT study found that the scoring models used in that tuning leaned left even when trained only to reward truthful answers. David Rozado's 2024 study found most chat models answered political questionnaires left of center while the base models beneath them did not, though he called the base-model result inconclusive. In a 2025 Stanford and Dartmouth study, about 10,000 Americans of both parties judged nearly all of 24 leading models left-leaning. And in 2024 Google said its tuning of Gemini's image generator "failed to account for cases that should clearly not show a range" of people, an episode Trump's executive order alludes to.
The same lean shows up in what chatbots cite. In NewsGuard's September audit the chatbots were three times as likely to cite the left-leaning look-alike sites as the right-leaning ones, and Gemini and Grok never cited a right-leaning one, though the analysts suggested the progressive sites' far more frequent posting may explain the gap better than any bias in the chatbots. Profoundd's own preliminary test, which judged identical claims with only the party's name swapped, found that the small open model Hermes 3, and two versions Profoundd trained from it, judged claims critical of Democrats more harshly by about a fifth of a step, a lean that came from Hermes 3, not from Profoundd's training. It found no average party lean in two Claude models, though under Profoundd's fair-reading rules single Claude verdicts moved with the party on 11 of 38 claims, eight of them in Democrats' favor, and Claude Sonnet ruled a claim more harshly when it named the party the claim is not usually made about. In NewsGuard's audit, Claude cited partisan look-alike sites in half of its answers.
What none of this shows is the rest of the reply's claim. These studies measure which way answers lean; none shows that anyone designed the lean or built models to lie, and the Stanford-Dartmouth study found that asking models to take a neutral stance made users rate their answers as more neutral. One point of context on this case: ChatGPT's praise of McCann did not come from its training. It came from pages an outside network planted on the open web, which a chatbot that searches the web will repeat whichever party they favor.
What Trump's order covers
Executive Order 14319, "Preventing Woke AI in the Federal Government," signed July 23, 2025, lets federal agencies buy only large language models that follow two principles: truth-seeking, which includes acknowledging uncertainty when reliable information is incomplete, and ideological neutrality. The White House budget office told agencies in December 2025 to write those terms into new contracts, and since August 2025 agencies have been able to buy ChatGPT Enterprise for $1 under a government-wide deal. So the order touches this case only indirectly, through the versions agencies buy. As Frye says, it does not solve the problem: it does not reach the consumer chatbots voters use, how they weigh an anonymous site that calls itself news, or who must disclose paying for political websites.
Questions that remain
Profoundd has drafted these questions; they have not yet been sent.
- For whoever runs the network: who paid, who chose the candidates, who or what wrote the articles, and whether any campaign or committee knew.
- For the State + Local Election Alliance: whether it, or anyone working for it, built, paid for or supplied material to any of the sites, and whether it knew its picks were being used.
- For Cloudflare and Netlify: what their rules require of anonymous political sites, and whether they will act on reports about these.
- For the 11 campaigns and the super PACs backing them: whether they or their vendors built, paid for or supplied material to any of the sites.
- For OpenAI: what changed after NPR flagged the sites, and whether ChatGPT checks who owns a site before citing it as news.
- For Google: which of its policies the sites broke, and when it acted.
- For the Federal Election Commission: whether any current rule requires sites like these to say who paid.
Disclosure
Profoundd's own robots.txt invites AI crawlers and says the site "exists to be cited in AI answers," and Profoundd publishes an llms.txt file. What NPR and VisionHeight flagged is the combination: sites that pose as local newsrooms, hide who runs them and are tuned for chatbots. Profoundd uses Claude, made by Anthropic, a competitor of OpenAI and Google, to research and draft its stories, as noted above.
Profoundd's read
Frye's post holds up as far as Profoundd could read it. Each fact in it matches NPR's report and the records Profoundd checked, and it stops where the evidence stops: it asks who paid instead of naming someone. Some coverage went further. Hot Air's headline called the network a "Democrat Scheme," while its text said no party organization is likely involved and concluded that "dark money" made it happen. The sites promote Democrats, and no one has publicly identified who built them. The strongest public lead is the candidate lists, which the statewide sites took from the State + Local Election Alliance's guide. That is a question for the alliance, not an answer.
The reply points at a real, measured lean in AI models but goes beyond the evidence in calling it a design to lie. This case also shows a weakness that does not depend on training: chatbots that search the web repeat what they find. An unknown operator, with what VisionHeight describes as one evening of setup and two weeks of computer-generated writing, got ChatGPT to cite its pages in 19 of NPR's 37 questions, each asked once. VisionHeight notes that any campaign, cause or foreign government could do the same. Its advice to voters is the practical part: ask the chatbot where its answer came from, and if the source is a site with no named owner or reporter, treat it "like an unsigned flyer."
Sources
- NPR, Huo Jingnan, "Websites pretending to be local news are boosting Democratic candidates," Oct. 30, 2026: npr.org
- VisionHeight, "Someone Is Manipulating Your AI to Swing Your Vote": visionheight.com
- Registration records, retrieved Nov. 2, 2026, from Verisign's .com registry (rdap.verisign.com) and Cloudflare's registrar (rdap.cloudflare.com), for all 37 domains; each site's robots.txt, llms.txt and home page, fetched the same morning. The network's addresses are listed in VisionHeight's report and are not linked here.
- State + Local Election Alliance voter guide, Mississippi page: localvoterguide.com; the alliance's website: state-elections.org; FEC committee C00812495: fec.gov, and the FEC's 2025-26 committee and operating-expenditure bulk files: fec.gov
- The Daily Reflector, Ginger Livingston, "PACs campaign for local Dems in county commission races," Oct. 28, 2026: reflector.com
- New York Post, Ariel Zilber, Oct. 30, 2026, McCann campaign statement: via Yahoo News
- Hot Air, David Strom, Oct. 31, 2026: hotair.com
- Politico Magazine, Ines Chomnalez and Isis Blachez of NewsGuard, "'Pink Slime' Is Infecting AI Chatbots Ahead of the Midterms," Sept. 25, 2026: politico.com (read via Yahoo's syndicated copy)
- NewsGuard, 2022 count of partisan local-news sites: newsguardtech.com
- Columbia Journalism Review, Tow Center, Priyanjana Bengani, on the money behind Metric Media, Aug. 8, 2024: cjr.org
- Federal Election Commission, advertising and disclaimers: fec.gov; 11 CFR 100.26: law.cornell.edu
- Executive Order 14319, Federal Register, July 28, 2025: govinfo.gov; OMB Memorandum M-26-04, Dec. 11, 2025: whitehouse.gov; GSA, ChatGPT Enterprise for agencies, Aug. 6, 2025: gsa.gov
- Feng, Park, Liu and Tsvetkov, "From Pretraining Data to Language Models to Downstream Tasks," ACL 2023: aclanthology.org
- Santurkar and colleagues, "Whose Opinions Do Language Models Reflect?," ICML 2023: mlr.press
- Fulay and colleagues, "On the Relationship between Truth and Political Bias in Language Models," EMNLP 2024: aclanthology.org
- David Rozado, "The political preferences of LLMs," PLOS ONE, July 31, 2024: plos.org
- Westwood, Grimmer and Hall, "Measuring Perceived Slant in Large Language Models Through User Evaluations," May 2025: stanford.edu
- Google, Prabhakar Raghavan, on Gemini's image generation, Feb. 23, 2024: blog.google
- The Washington Post's 2023 analysis of the C4 training set, as reported by The Register: theregister.com
- Laubacher as Colorado's 4th District Democratic nominee: Axios
- Profoundd, "Testing Our Own AI for Political Bias: A Preliminary Report": profoundd.com
