AI chatbots are becoming a default layer of political communication, but they do not produce the same reality in every language. EU-funded research auditing six major LLMs shows measurable narrative divergence: the same prompt can produce accurate answers in English and propaganda-aligned outputs in Russian. Leading Western AI models have a significant risk: false balance, framing verified facts as competing narratives. This is a communication risk. If your message travels across languages, you may already be losing control over how it is interpreted. This session presents real audit evidence, demonstrates how divergence appears in practice, and outlines how political actors can start testing and monitoring these risks.