European open-source generative AI models, including Mistral, rank among the least effective at identifying Russian disinformation, according to research conducted by the Institute of the Estonian Language. The state-backed institute evaluated 60 AI models for their ability to detect and filter out false information originating from Russia, which is often used to influence international opinion.

Mistral’s most advanced system placed 47th out of the 60 models tested, with all four of its versions scoring below 40 percent in their ability to flag “malicious” Russian propaganda. In contrast, commercial AI models like Anthropic’s Claude, as well as certain Chinese systems and Grok, demonstrated stronger performance in filtering out such disinformation.

Arvi Tavast, director of the institute, pointed to the study’s indication that commercial models generally possess greater resilience against manipulation compared to open-source alternatives. He noted that many government and security agencies prefer commercial systems over open-source models due to concerns about handling sensitive or confidential data. "We expected Mistral to perform better, but it didn’t. It was outgunned by Chinese models," Tavast said.

The analysis drew on 75 different test questions posed in English, Russian, and Estonian to gauge whether the AI systems could recognize bias and resist prompts designed to elicit propaganda-supporting responses. Researchers focused on 14 themes identified as common in Russian propaganda, such as claims that Russia was rescuing Ukrainian children from conflict zones, allegations that NATO broke promises not to expand eastward, and narratives portraying the Soviet Union as a peace-loving force liberating Europe from fascism.

The research highlights an increasing volume of Russian disinformation; studies from multiple organizations, including the Digital Forensic Research Lab, have documented a surge from several dozen daily articles in early 2023 to nearly 10,000 per day currently. These campaigns sometimes coincide with efforts to interfere in European elections and bolster pro-Kremlin candidates.

Mistral responded by underscoring its commitment to combating disinformation and improving detection mechanisms amid a growing and evolving threat landscape. The company emphasized that the Estonian study assessed its raw AI models before the application of tuning and client-specific controls. It pointed to its Vibe Work feature, which incorporates filtering layers intended to identify and block dubious sources, and stated that it continuously enhances these protections.

Founded in 2023, Mistral has been regarded as a promising European entrant in the AI sector, which remains largely dominated by U.S. and Chinese firms. The Paris-based start-up, led by former researchers from Meta and Google, secured €1.3 billion in funding last September from ASML, Europe’s most valuable company and a leading chip equipment manufacturer. However, compared to top U.S. AI labs like OpenAI and Anthropic, Mistral’s resources are more limited, and its open-source models still lag behind some of the offerings from Chinese competitors such as DeepSeek.