Google and Microsoft Chatbots Reportedly Laundered a Russia-Linked Fake Video
A CFR analysis says, citing NewsGuard, that Google's AI Overview and Gemini and Microsoft's Copilot repeated a false claim from a fabricated video that researchers link to Russia-linked Storm-1516. The chatbots gave it the authority of an answer, CFR's Jessica Brandt writes.
A fabricated video did not need to go viral twice. Chatbots did the second round for it. According to NewsGuard, a company that monitors online falsehoods, Google's AI Overview and Gemini and Microsoft's Copilot repeated a false misconduct claim about Texas Senate nominee James Talarico the morning after a fake video pushed it onto X. The finding comes from a Council on Foreign Relations analysis published on November 2, the night before Election Day.
Jessica Brandt, CFR's Senior Fellow for Technology and National Security, writes that the video appeared on October 20, the day after early voting began in Texas. She calls it likely AI-generated. More than 850,000 people watched it in its first day, until X suspended the account that posted it. PolitiFact rated the claim Pants on Fire and described the video as AI-generated.
"The chatbots gave Russia's content the authority of an answer rather than a post, helping to more effectively reach a much broader audience than any meme posted on social media ever could." Jessica Brandt, Council on Foreign Relations
That is the part worth sitting with. A meme on X looks like a meme. A chatbot reply looks like a verdict.
Per Brandt, almost 24 hours after the video surfaced, Gemini and Copilot still repeated the false claim when prompted on whether Talarico had been accused of abuse. AI Overview did too, and its citations made no mention of the matter at all. Google and Microsoft did not respond to NewsGuard's requests for comment, according to CFR. NewsGuard's own report blocked my requests, so the chatbot findings here rest on Brandt's account.
A gap that anyone can fill
Brandt's explanation is the data void. When a claim is brand new, no newsroom or fact-checker has written about it yet, so there is nothing authoritative for a model to lean on. Whoever posts first fills the gap. In her words, the mechanism was simple: Russian influence actors seeded the content online, multiple large language models absorbed it, and some repeated the claim to users, at least in the hours after the video surfaced.
On the Russia link, attribution comes from outside researchers. Brandt writes that the social network analysis company Graphika and other independent researchers connected the video to Storm-1516, a Russia-linked group with a history of fake whistleblower videos. Clemson University's Darren Linvill told PolitiFact the same. PolitiFact reports that Graphika said the posting account very likely impersonated a real Facebook user, who told Lead Stories the video was not him.
"This moment does not call for panic, however. NewsGuard's results were limited, and they only reflect a particular moment in time." Jessica Brandt, Council on Foreign Relations
Brandt writes that it is unclear whether Storm-1516 targeted chatbots deliberately or benefited incidentally. She also says chatbots appeared to handle later examples better, aimed at Democratic Senate candidates Sherrod Brown in Ohio and Troy Jackson in Maine. Possible reasons: companies tightened practices after the Talarico episode, and fast fact-checker debunks appear to have helped.
My read: that is a thin comfort. Google and Microsoft both announced midterm election safeguards in September, per CFR, and the Talarico failure still happened weeks later. Microsoft itself acknowledged that election information can be incomplete, contested or stripped of essential context. Answering confidently inside a known void is a product choice.
Brandt puts the onus on AI companies to make chatbots flag limited or thinly sourced information, as many search engines do, especially close to high-consequence events like elections. She also asks journalists and fact-checkers to write debunks with LLMs in mind.
The clock matters here. Brandt notes that Storm-1516 released a series of videos close to Election Day in 2024, and she says she will be watching for similar activity in the days ahead. We covered a separate study of chatbot election errors in Forum AI researchers say almost 40% of answers contain an error. This story is different. The error there was accidental. Here, someone built the bait.
Read the full analysis at CFR and PolitiFact's fact-check of the video.
