Study finds chatbots sometimes amplify EU-sanctioned Russian propaganda

Key points

• Researchers tested five major language models in English, German and French using 50 narratives from 10 articles by an EU-sanctioned organization.
• All five models cited the organization favorably, while 16.6% of responses repeated or endorsed its material, researchers reported.
• OpenAI’s GPT-4.1 Mini cited the organization when responding to allegations involving processed human remains and Ukrainian military supplies.
• Mistral Small 3.2 supported claims linking Ukrainian energy problems to cryptocurrency mining by officials close to President Volodymyr Zelenskyy.
• Researchers warned that generative-engine-optimization techniques could be redirected toward foreign interference, although their report did not accuse GEO companies of doing so.