Pasi Eronen Watt 🇺🇦

🤖🇪🇪 How resistant is AI to Kremlin propaganda? Together with the Institute of the Estonian Language, Propastop helped develop a new propaganda resilience benchmark that tested more than 50 language models. The results were revealing. While the strongest models generally recognized manipulative narratives, performance varied widely. Some models were significantly more vulnerable when questions contained misleading assumptions or were intentionally designed to push them toward a propaganda-friendly answer. In a few cases, the probability of producing a propaganda-supporting response nearly doubled. 📈 The benchmark was built around 14 key narratives used in Russian influence operations and included 75 questions in Estonian, English, and Russian. Models were tested not only on what they know, but on whether they can recognize manipulation when it is hidden inside the question itself. 🎭 As AI becomes a major source of information for millions of people, its resilience to propaganda is no longer just a technical issue. It is also a question of information security. 🔍🛡️ The full benchmark is available at xn--mdupuu-pxaa.eki.ee.