The European AI model Mistral has proven vulnerable to Russian propaganda, according to the Financial Times
A popular artificial intelligence model developed by the French startup Mistral has demonstrated a low ability to detect and filter Russian disinformation. These conclusions were reached by researchers at the Institute of the Estonian Language, who analyzed the effectiveness of dozens of modern generative AI systems.
According to the study’s findings, open-source models proved to be the most vulnerable to the influence of Russian propaganda and manipulative narratives.
These are systems whose architecture and source code are available for public use and modification.
As noted, Mistral ranked only 47th out of 60 tested AI models. All four tested versions of the French system scored below 40% on the metric for identifying sources that researchers classified as disseminators of harmful Russian propaganda.
Analysts concluded that Mistral lags significantly behind its competitors in its ability to recognize false or manipulative statements used by Russia to influence an international audience.
At the same time, proprietary commercial models demonstrated better results. In particular, the following were more effective at detecting propaganda narratives:
ChatGPT;
Claude by Anthropic;
Grok by xAI;
certain Chinese AI systems.
Researchers emphasize that the challenge of countering disinformation is becoming increasingly urgent amid the rapid proliferation of generative AI. The ability of models to detect fakes, propaganda, and manipulative content is increasingly seen as one of the key criteria for assessing their reliability and safety.
The Financial Times reports on this.
French AI startup Mistral AI raised €385 million in a funding round. As a result, the estimated valuation of the startup, which was founded less than a year ago, has risen to approximately $2 billion.
