The rapid rise of generative artificial intelligence has brought with it a significant, albeit quiet, battleground: the fight against state-sponsored misinformation. As global political tensions simmer, researchers are increasingly looking at how these powerful digital tools—which we rely on for information, analysis, and education—handle the surge of propaganda flooding the internet. A striking new study from the Institute of the Estonian Language serves as a wake-up call, revealing that not all AI models are created equal when it comes to integrity. By testing 60 different systems against a battery of questions rooted in known Russian propaganda, the research found that some of our most celebrated models, particularly in the open-source sector, are surprisingly vulnerable to manipulation.
The study’s most provocative finding involves the European AI champion, Mistral. While the company is often hailed as a symbol of European technological sovereignty, its models performed significantly worse than expected, ranking 47th out of 60. Even more surprising was the fact that these models were outperformed by various Chinese systems, as well as mainstream commercial giants like Anthropic’s Claude. This discrepancy suggests that the “open-source” nature of these tools, while beneficial for accessibility and local customization, may also leave the door slightly ajar for bad actors. As Arvi Tavast, the institute’s director, bluntly noted, the ease of deploying these open models may come at the cost of the robust safety guardrails that deep-pocketed commercial firms have spent years refining.
To reach these conclusions, the Estonian researchers challenged the models with 75 distinct questions across English, Russian, and Estonian. They focused on 14 specific, well-documented themes of Russian propaganda, such as the revisionist claims that the USSR was merely a “peace-loving liberator” of Europe or that the current invasion of Ukraine is actually a humanitarian evacuation of children. When quizzed on these topics, many models struggled to identify the bias or, worse, echoed the narratives intended to confuse the public. This is a critical concern, especially as governments—including Estonia’s—incorporate these very tools into school systems to help students learn and conduct research.
In response, Mistral has pushed back, arguing that the Estonian team tested their “raw” or baseline models rather than the fully optimized versions that customers actually use. The company emphasized that its products are designed with layered filtering to block deceptive content and that they are continuously upgrading these safeguards to keep pace with the evolving threat of misinformation. This tension highlights an ongoing industry debate: should we prioritize the open, transparent, and customizable nature of AI that allows nations to maintain control over their data, or should we favor the more locked-down, heavily moderated models offered by the tech giants of Silicon Valley?
The implications of this study reach far beyond software testing; they strike at the heart of our digital security. With organizations like the Digital Forensic Research Lab reporting that Russian propaganda has jumped from a few dozen articles a day to nearly 10,000, AI systems are no longer just passive observers—they are potential conduits for influence operations. If an AI can be coaxed into validating false historical narratives or legitimizing election interference, it becomes a weapon of mass persuasion. Governments are now forced to weigh the benefit of using “sovereign” European technology against the stark reality that safety and security are currently skewed toward established, commercial American and Chinese players.
Ultimately, this research serves as a sobering reminder that innovation must be tempered by responsibility. European companies like Mistral have ambitious plans to invest billions in data centers and infrastructure to rival the dominance of US Big Tech, and for good reason—nations need reliable, independent alternatives to ensure their data remains secure. However, as these tools become deeply embedded in our schools, boardrooms, and democratic processes, their “moral compass”—or their ability to discern fact from state-funded fiction—will be the most important feature they possess. Bridging the gap between the speed of deployment and the necessity of truth remains the greatest challenge for the next generation of AI developers.