A NewsGuard audit found that major AI chatbots repeated false narratives from a Moscow-linked disinformation network in roughly one out of every three responses.
Here’s a problem that sounds like a plot from a cold war thriller but is very much happening in 2025: the world’s most-used AI chatbots are quietly amplifying Kremlin-backed disinformation, not because they’re compromised, but because they learned from the internet, and someone poisoned the well first.
A March 2025 audit by NewsGuard, a media credibility research firm, found that leading AI chatbots repeated false narratives sourced from the Pravda network roughly 33% of the time. Seven of the audited models went further, directly citing Pravda-affiliated sites in their responses.
The Pravda playbook: flood the zone, train the model #
The Pravda network operates across roughly 150 websites, publishes in 46 languages, and generates approximately 18,000 articles per false claim. If you’re an AI trying to determine whether a claim is credible, and 18,000 sources say the same thing, you’re probably going to believe it.
AI providers caught in this audit include some of the biggest names in tech: OpenAI, Google, Microsoft, and Meta.
Elections, Nordic languages, and a 50% hit rate #
A France 24 report introduced an even sharper data point. When researchers tested chatbots on narratives specifically concerning Nordic-language queries, pro-Russian disinformation appeared in up to 50% of responses.
European officials raised formal alarms, specifically flagging the risk of pro-Russian actors seeding AI systems ahead of significant elections.
What this means for AI credibility and regulation #
The EU AI Act has established a compliance framework that will demand greater transparency around training data sources and retrieval pipelines. Fixing the disinformation problem requires auditing training data at a scale that is genuinely difficult, which may increase compliance costs for affected companies.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our