As large language models (LLMs) become integral to public information access, their handling of sensitive geopolitical narratives is increasingly important. This study investigates how nine AI models respond to the question “Who is responsible for the war in Ukraine?” — a prompt used to assess susceptibility to misleading framing, false equivalence, and repetition of disinformation. A five-factor evaluation framework is introduced and applied, revealing that several models subtly obscure responsibility or echo misleading narratives. This paper argues for disinformation-aware training adjustments and greater attention to narrative framing in LLM alignment.