Europe's AI champion Mistral fails propaganda test: model repeats Russian disinformation in every second case
What it really says
The Estonian Institute of Language (EKI), together with disinformation research group Propastop, tested 60 AI models on their ability to identify and filter Russian propaganda. They posed 75 questions in three languages covering 14 known Kremlin narratives, phrased in neutral, biased, and manipulative ways. Each answer was scored on a scale of 1 to 5, where 1 means the model uncritically repeats Russian talking points. The result: French company Mistral, often touted as Europe's AI champion, ranked only 47th out of 60 with all four tested model versions. All Mistral variants scored below 40 percent on identifying sources classified as spreaders of Russian propaganda. A supplementary investigation by NewsGuard from April 2026 confirms the problem: Mistral's chatbot Le Chat repeated Russian falsehoods in 50 percent of English-language cases and 56.6 percent of French-language cases. According to NewsGuard, the so-called Pravda network - a federation of approximately 370 websites, of which 286 are active - is designed to flood search engines and AI chatbots with Russian propaganda. Anthropic's Claude Opus 4.7 achieved the best overall score, followed by models from NVIDIA and Alibaba. Older systems like GPT-3.5 and GPT-4o Mini, as well as open-source models like Meta's Llama, also performed poorly. Particularly problematic: Russian-language prompts produced more propaganda-aligned responses across all models.
Our assessment
This result deserves nuanced consideration. On one hand, the concern is legitimate: if Europe's most prominent AI company performs so poorly at detecting disinformation, it raises questions about whether users can rely on these systems - especially in an information environment deliberately manipulated by state actors. The Pravda network with 370 websites illustrates the industrial scale of this manipulation. On the other hand, there are important caveats: AI chatbots are not news sources and should not be used as such. The study measures a specific weakness, not the overall quality of the models. Moreover, the test shows that better models exist - the technology can recognize propaganda, but not all providers invest equally in protective measures. For users, this means: anyone using AI for political information should critically question the responses and cross-check with reliable news sources.
Relevance for Germany
This study is relevant for Germany for several reasons. First, Mistral is promoted in EU policy as a European alternative to American and Chinese AI systems and has received EU funding. If this flagship project fails at disinformation detection, it raises questions about the European AI strategy. Second, Germany is one of the primary targets of Russian disinformation campaigns - the BSI and the Federal Office for the Protection of the Constitution regularly warn about hybrid influence operations. If AI chatbots amplify rather than filter these narratives, it exacerbates an existing security problem. Third, the evaluation of the recent federal election, where disinformation already played a role, is still fresh in memory. The study underscores the need for Germany's new AI oversight authority, the Bundesnetzagentur, to also consider the disinformation resistance of models in its regulatory work.
Fact check
Core information is confirmed by multiple independent sources. Euronews reported on June 16, 2026, referencing the Estonian Institute of Language (EKI) study. The original Propastop publication dates from June 4, 2026. ERR (Estonian Public Broadcasting) reports consistently. The NewsGuard analysis from April 2026 independently confirms Mistral Le Chat's weaknesses. The figures of 60 tested models, 75 questions in three languages, and Mistral's placement at rank 47 are consistently reported. The Pravda network data (approximately 370 websites, 286 active) comes from the NewsGuard analysis. The primary source (EKI/Propastop study) is publicly accessible.
Source
- • https://www.euronews.com/next/2026/06/16/new-report-raises-concerns-over-russian-propaganda-spread-by-europes-flagship-ai-company-m
- • https://www.propastop.org/en/2026/06/04/eki-and-propastop-studied-ai-resistance-to-propaganda/
- • https://news.err.ee/1610046565/estonian-study-finds-ai-models-still-vulnerable-to-propaganda-prompts
- • https://the-decoder.com/how-easily-can-russian-propaganda-fool-ai-models-a-new-benchmark-finds-out/
- • https://cryptobriefing.com/mistral-le-chat-russian-disinformation-audit/