A network of websites masquerading as independent local news outlets is systematically targeting competitive U.S. House races with content favourable to Democratic candidates, new research shows. The operation, whose backers remain undisclosed, has managed to infiltrate the knowledge base of OpenAI’s ChatGPT while largely evading Google’s search and news algorithms.
The discovery raises urgent questions about the integrity of the information ecosystem ahead of the November elections. As voters increasingly rely on AI-generated summaries for civic information, the ability of partisan actors to seed those systems with disguised political messaging represents a novel threat — one that traditional search safeguards appear better equipped to handle than the new generation of conversational agents.
The anatomy of a disguised network
Researchers tracking the cluster describe a sophisticated mimicry of legitimate local journalism. The sites adopt mastheads that evoke community newspapers — names incorporating “Gazette,” “Times,” “Post,” and “Chronicle” alongside geographic identifiers for specific suburbs, counties, and congressional districts. Their layouts ape the clean, ad-supported templates of genuine regional publishers. But the bylines are often fictitious, the mastheads list no physical address, and the ownership trails dissolve into privacy-protected domain registrations.
What distinguishes this network from the “pink slime” operations documented by organisations such as NewsGuard and the Tow Center for Digital Journalism is its precision targeting. Rather than blanketing entire states with low-effort content farms, these outlets concentrate their output on a handful of toss-up House districts. Articles highlight Democratic incumbents’ constituent service, amplify campaign talking points on healthcare and infrastructure, and frame Republican challengers in uniformly negative terms — all presented with the ostensible neutrality of straight news reporting.
The volume is modest by content-farm standards: dozens of articles per site per month, not hundreds. That restraint appears deliberate, designed to evade the spam classifiers that flag high-velocity, low-quality publishers. It also mirrors the publication cadence of authentic weekly or bi-weekly community papers, reinforcing the illusion of legitimacy.
AI systems diverge on handling partisan mimics
The most consequential finding concerns how the major AI platforms treat this material. When researchers queried ChatGPT about candidates and races in the targeted districts, the model repeatedly surfaced summaries drawn from the disguised sites, presenting their framing as factual context. In several instances, ChatGPT cited the outlets by name, attributing political claims to “local news reports” without disclosing their partisan provenance.
Google’s Search and News products, by contrast, returned almost no visibility for the same domains. The company’s long-standing investments in publisher verification, E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) signals, and manual quality rater guidelines appear to have effectively quarantined the network. A Google spokesperson declined to comment on specific domains but noted that the company’s systems are designed to “prioritise original, trustworthy reporting from established journalistic institutions.”
The divergence underscores a growing asymmetry. Traditional search engines have spent two decades building — and constantly refining — defences against manipulation. Large language models, trained on static snapshots of the web and fine-tuned with reinforcement learning from human feedback, inherit the biases and gaps of their training data. If a partisan network achieves sufficient scale and semantic coherence before a model’s knowledge cutoff, its perspective can become baked into the model’s representation of reality.
The transparency vacuum
No organisation or individual has claimed responsibility for the network. Domain registration records are shielded by privacy services. The sites carry no “About Us” pages disclosing funding, editorial independence policies, or correction procedures. Their social media amplification follows a pattern consistent with coordinated inauthentic behaviour: clusters of accounts with limited posting histories sharing each article within minutes of publication, often using identical or near-identical language.
Campaign finance filings offer no obvious trail. The sites do not appear as expenditures for any candidate committee, party committee, or registered super PAC. That absence could indicate funding through a 501(c)(4) “dark money” vehicle, which is not required to disclose donors, or through a limited liability company structured to avoid political committee registration thresholds.
Legal experts note that the operation likely skirts existing disclosure requirements. The Federal Election Commission’s rules on “coordinated communications” and “independent expenditures” were written for broadcast, mail, and digital advertising — not for synthetic news outlets that publish ostensibly editorial content. The Honest Ads Act, which would extend disclosure to online political advertising, has stalled in Congress. The DISCLOSE Act, targeting dark money, has met similar resistance.
Why it Matters
The emergence of partisan actors weaponising the credibility of local journalism — and successfully poisoning the well of AI-generated civic information — represents a qualitative escalation in digital political manipulation. Voters in competitive districts deserve to know when the “local news” informing their choices is actually a campaign brochure in disguise. Until platform governance and regulatory frameworks catch up to this tactic, the burden falls on news literacy: readers must verify the provenance of every outlet that shapes their understanding of the ballot.
