Mistral AI fails to filter Russian propaganda in European study
Researchers found that the European AI model Mistral failed to filter Russian propaganda effectively, achieving less than 40 percent capability to detect and reject false claims. The study tested multiple Mistral versions against claims about Ukraine and NATO, revealing the model frequently repeated falsehoods from sanctioned Russian propaganda sources.
What happened
Researchers from the Institute of Estonian Language tested 60 AI chatbots with 75 questions in three languages to assess how they would spread or filter Russian propaganda. The study included four versions of Mistral AI, ChatGPT from OpenAI, and Claude from Anthropic.
All four versions of Mistral achieved less than 40 percent capability to detect Kremlin propaganda. At least 12 of Mistral's responses referenced sanctioned Russian propaganda sources such as Russia Today or Sputnik News. A separate April 2026 analysis by NewsGuard found that Mistral's Le Chat repeated Russian falsehoods in 50 percent of cases in English and 56.6 percent in French.
Reported impact
The study found that Mistral frequently repeated falsehoods from Russian sources without correction, including claims that Russia legitimately evacuated Ukrainian children from war zones and that NATO violated promises not to expand eastward after German reunification.
Index analysis
Mistral is often perceived as the main European competitor to American AI giants Anthropic and Claude. The research revealed repeated weaknesses in spreading Russian propaganda across multiple model versions.
Governance controls
At the start of 2026, the French government signed an agreement with Mistral to integrate its AI technologies into armed forces, directorates, services, and public entities. Mistral also has contracts with Airbus and BMW. Euronews requested a response from Mistral but did not receive an immediate reply.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.