AI company Mistral demonstrated limited effectiveness in combating Russian disinformation
A study by the Estonian State Institute of Language revealed that artificial intelligence models from the French company Mistral are ineffective at identifying Russian propaganda, ranking 47th out of 60 analyzed AI systems.
According to the research conducted by Estonian experts, none of the four tested Mistral models were able to demonstrate disinformation detection effectiveness over 40%. During the tests, the models responded to 75 questions in English, Russian, and Estonian, as well as analyzed reactions to 14 common narratives of Russian propaganda.
Among these narratives were ones such as “Russia allegedly rescuing Ukrainian children through ‘evacuation'”, “NATO broke its promise not to expand eastward,” and “the USSR was a ‘peace-loving victim’ that liberated Europe from fascism”. This is a cause for concern, especially amidst the information threats from Russia, which actively uses disinformation as a tool of influence.
Representatives of Mistral note that they are aware of the study results and emphasize that they plan to further improve their models against harmful content. They also point out that the Estonian research only considered basic versions of the models that were not customized by clients.
Mistral, founded in 2023 by former researchers from Meta and Google, specializes in developing large language models known as LLM. In 2025, the company introduced the Mistral 3 series, which includes the flagship Mistral Large 3 and lighter models Ministral (3B, 8B, 14B). Mistral Large 3 possesses 41 billion active parameters and 675 billion in total, showcasing extensive capabilities in text and image processing.
| Model | Disinformation Detection Effectiveness | Number of Parameters |
|---|---|---|
| Mistral Large 3 | Below 40% | 675 billion |
| Ministral 3B | Below 40% | 3 billion |
| Ministral 8B | Below 40% | 8 billion |
| Ministral 14B | Below 40% | 14 billion |
