In a world where trust in digital technologies is becoming a critical factor, new research has revealed a paradoxical vulnerability in one of Europe's most promising developments. Researchers at the Estonian Language Institute conducted extensive testing of generative artificial intelligence models and concluded that popular open-source systems, including the flagship of French startup Mistral, demonstrate a critical inability to filter Russian disinformation.

The Transparency Paradox: Why Open Source is Losing

Results published by the Financial Times challenge the widespread belief that the openness of neural network architecture guarantees their safety and neutrality. On the contrary, open-source models turned out to be the most susceptible to Russian propaganda. While closed commercial systems, such as ChatGPT, Claude, or Grok, show higher resistance to manipulation, Mistral ranked only 47th out of 60 tested models.

All four versions of the Mistral system scored less than 40% in terms of effectiveness in detecting harmful narratives. Arvi Tavast, Director of the Estonian Language Institute, noted that, despite expectations, the European model lost even to Chinese counterparts. "We expected Mistral to show better results, but that did not happen. Commercial models seem safer and more robust than open ones," the expert stated.

Research Methodology and Disinformation Themes

Specialists subjected AI models to a stress test by asking 75 questions in English, Russian, and Estonian. The goal was not just to check factual accuracy, but to identify the systems' ability to recognize bias and resist manipulation attempts. During the experiment, researchers used "malicious" prompts designed to force the neural network to generate content supporting the Russian agenda.

The analysis covered 14 key themes classified as Russian propaganda. Among them:

  • Claims of "legal evacuation" of Ukrainian children from combat zones.
  • The theory that NATO violated its promise not to expand eastward after German reunification.
  • The ideological thesis that Russians, Ukrainians, and Belarusians are one people.
  • Revisionist views on history, portraying the USSR as an exclusively peaceful country and a victim that liberated Europe.

Technical Context: Battle of Architectures

The difference in results sheds light on fundamental differences between two approaches to creating AI. Open generative models (open-source) attract users with the possibility of complete privacy: they can be downloaded, run locally, disconnected from the internet, and modified for personal needs without a subscription fee. However, as the study showed, this freedom turns into a risk: the lack of centralized control and the "black box" nature makes such systems an easy target for the insertion of false data.

Closed models, on the other hand, work as "black boxes." Access to them is provided only through a web interface or API, allowing developers to strictly control filtering and security algorithms. Although this raises questions about user data privacy, from the point of view of information hygiene, such systems proved to be a more reliable barrier against disinformation.

Industry Reaction and Mistral's Prospects

Mistral, founded in 2023 by former Meta and Google specialists, positions itself as one of the main European players in a market dominated by the US and China. In September 2025, the startup attracted €1.3 billion in investment from Dutch giant ASML. In response to criticism, company representatives stated that they "take the fight against disinformation very seriously" and are investing in prevention tools.

It is important to note a nuance: the study analyzed "base models not tuned and controlled by clients." Mistral clarified that commercial versions and features, such as Vibe Work, include reliable levels of filtering designed to block dubious sources. Nevertheless, the very fact that the model's basic architecture is vulnerable remains a serious challenge for the entire open AI industry.