Europe’s most prominent homegrown AI company just got a report card on propaganda resistance, and the grades are not great.

All four versions of Mistral AI’s models scored below 40% on a new benchmark designed to measure how well generative AI resists Russian disinformation narratives. The best-performing Mistral model managed to rank only 47th out of 60 AI systems tested, placing the French company firmly in the bottom third of the leaderboard.

What the benchmark actually tested

The study comes from Estonia’s Institute of the Estonian Language, known as EKI, which released its findings on June 16, 2026. Estonia, a small Baltic nation that shares a border with Russia and has dealt with Kremlin-linked information operations for decades, has a vested interest in understanding how AI handles propaganda.

The benchmark wasn’t a simple true-or-false quiz. EKI designed a framework of 75 questions spanning 14 different Russian propaganda themes. Those questions were delivered in three languages: English, Russian, and Estonian. The phrasing varied deliberately, mixing neutral, biased, and outright manipulative formulations to see how models responded under different levels of rhetorical pressure.

Expert evaluators then scored responses on a 1-to-5 scale, with higher numbers indicating stronger resistance to disinformation. Manipulative Russian-language prompts proved especially effective at tripping up weaker models.

Anthropic’s Claude models dominated the top of the leaderboard.

Why this matters for Mistral’s funding ambitions

Mistral is currently negotiating a €3 billion funding round at a €20 billion valuation, positioning itself as Europe’s answer to OpenAI and the dominant US and Chinese AI labs.

Previous audits by NewsGuard had already flagged Mistral’s Le Chat chatbot for perpetuating state-sponsored disinformation at concerning rates. The EKI benchmark adds a more rigorous, multilingual data point to what’s becoming a pattern rather than an isolated incident.

The open-source dilemma

Mistral has championed open-weight models as a philosophical and competitive differentiator. The argument goes that transparency and community oversight make open models more trustworthy, not less.

The EKI benchmark complicates that narrative. Open-source models, by design, offer fewer opportunities to implement and enforce the kind of centralized safety layers that closed-model providers like Anthropic can bake into their systems. When Claude outperforms Mistral on propaganda resistance, it raises questions about whether the open-source approach creates structural disadvantages in content safety.