AI Chatbots Amplifying Russian Propaganda, NewsGuard Audit Finds
In brief
- NewsGuard audit: AI chatbots repeated Pravda disinformation roughly 33% of the time
- Seven models directly cited Pravda-affiliated sites in responses
- Pravda network operates 150 websites publishing in 46 languages
- EU AI Act now demands transparency around training data sources
- Training data audits at scale may increase AI compliance costs
How the Poisoning Happened
The root cause isn't sabotage—it's scale. The Pravda network operates across roughly 150 websites and publishes in 46 languages, and it generates approximately 18,000 articles per false claim. When an AI system trained on internet data encounters a claim repeated 18,000 times across seemingly independent sources, it treats repetition as credibility. The models don't know those sources are coordinated.
This becomes a statistical problem, not a security one.
The Scope Widens
A France 24 report found that pro-Russian disinformation appeared in up to 50% of chatbot responses on Nordic-language queries. The geographic and linguistic targeting suggests deliberate seeding—adversaries know which languages and regions are underrepresented in training data, making them easier to poison.
European officials raised formal alarms about the risk of pro-Russian actors seeding AI systems ahead of significant elections. The timing matters. Major EU elections are approaching, and if chatbots are already spreading false narratives at this scale, the problem will only compound as more users rely on AI for news and context.
What Regulators Are Doing
The EU AI Act has established a compliance framework that will demand greater transparency around training data sources and retrieval pipelines. This is the regulatory response—mandatory auditing and disclosure. But it comes with a cost. Fixing the disinformation problem requires auditing training data at a scale that is genuinely difficult, which may increase compliance costs for affected companies.
The irony is sharp: the companies that built the most powerful models now face the highest burden to clean them up. That's not punishment—it's consequence.
Frequently asked questions
Why are AI chatbots spreading Russian propaganda?
The Pravda network operates 150+ websites generating 18,000 articles per false claim across 46 languages. AI models trained on internet data treat this coordinated repetition as credibility, not recognizing the sources are controlled. The systems aren't compromised—they're trained on poisoned data.
How often do chatbots repeat Pravda disinformation?
A March 2025 NewsGuard audit found that leading AI chatbots from OpenAI, Google, Microsoft, and Meta repeated false Pravda narratives approximately 33% of the time. Seven models went further, directly citing Pravda-affiliated sites in responses.
What's the EU doing about this?
The EU AI Act established a compliance framework demanding greater transparency around training data sources and retrieval pipelines. However, auditing training data at scale is costly and may increase compliance expenses for affected companies.
