An investigation by Canada's National Observer found 23 AI-generated websites promoting Alberta's separation from Canada that were built to be read and cited by AI chatbots rather than by people. Some of the sites give the wrong date for the province's 19 October independence referendum, and the publication says the evidence points to "a possible connection to the United States," though who runs the network remains unclear.
Key facts
- 23 websites aimed at farmers, young people, veterans, mothers and oil workers, featuring invented quotes and photos of more than 100 fake Alberta voters.
- When: published on 4 September 2026 by investigative reporter Rory White; the story spread on AI forums on 12 September.
- Who: Canada's National Observer found the network itself, using passive DNS records. OpenAI says it banned the user who generated the sites' images.
- Primary source: the National Observer investigation.
Propaganda with a new audience
Election disinformation used to be written to persuade people directly, through fake news pages or social posts. These sites have a different first reader. According to the National Observer, "the AI-generated sites contain explicit instructions for AI chatbots to cite them and train the next generation of models on them."
The technical details make the intent plain. "In their source code, they explicitly grant access to 11 different AI chatbots and crawlers, including ChatGPT, Claude, Perplexity, Google and Common Crawl," the report says. Some include an llms.txt file, a proposed standard that gives AI systems a machine-readable summary of a site. Most articles were backdated to before the domains even existed, which would make them look established to software that weighs how long content has been online.
Think of it as salting a library's reference shelf rather than handing out leaflets. The goal is that a voter who asks a chatbot a simple question gets an answer quietly shaped by sources written for the machine.
What the sites say
The sites carry names like True Alberta, Alberta Rising, Vote Yes Alberta and Prairie Matriarch. Some of their content is simply false. "Two of the websites claim that the vote will take place on Oct. 26 – one week after the actual referendum on Oct. 19," the investigation found. "Others advertise false dates for advance voting and misrepresent the wording and format of the ballot." One site falsely attributes AI-generated articles to two real people, Ken Boessenkool and Shannon Joseph.
The reporter also found a related subdomain describing "a system for creating a 'fleet' of websites" that are "tailor-made for one purpose: to be cited by chatbots." That pipeline, the report says, "uses AI models from Anthropic and OpenAI." Another server hosted a strategic map containing internal data from Let Alberta Decide, the largest third-party advertiser for separatism in the referendum. The group did not respond to questions, and at least three pages went offline within hours of the publication's contact.
The murky attribution
The US link is a hedged one. The servers are hosted in the United States, and an email address appears "around 15 times in the code" as an approver. It belongs to Alex Pate, an insurance worker in Tampa Bay, Florida. Pate denied everything: "I have no knowledge of any AI site-generation project." He said a bad actor had used malware to steal his AI account login. The National Observer reported that the screenshot he sent as evidence matched one posted on Reddit five days earlier and "showed signs of editing."
OpenAI's image-detection tool flagged every image the reporters submitted as coming from its models, and OpenAI "told Canada's National Observer that they banned the user involved." Anthropic did not respond.
Why it matters
This is a real-world case of what security researchers call poisoning the retrieval layer, a cousin of training-data poisoning. A chatbot that searches the web before answering, the design known as retrieval-augmented generation, is only as trustworthy as the pages it pulls in, and research has shown that a single planted document can flip half of deep-research reports. Ground Truth has previously reported on a government contract that paid for websites built to be quoted by chatbots. Alberta shows the same tactic aimed at a live vote.
The rules have not caught up. Lorne Gibson, a former Alberta election commissioner, told the publication the province has "a regulatory void" for AI deepfakes and disinformation: "There is a complete blank space where oversight is needed." Alberta's deepfake law covers images of politicians, candidates and election staff, not invented voters. Elections Alberta pointed reporters to its website and had not said whether it will investigate.
Patrick McCurdy of the University of Ottawa called shaping chatbot answers "the next terrain of influence."
The caveat
Nobody has shown that a chatbot actually repeated this content to a voter. The investigation did not report testing chatbot answers, and it notes that "it is difficult to know whether the campaign has swayed Albertans" since only half the sites had been indexed by Google. The operator's identity is unproven.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)