Open-Weight LLMs Lead in Resisting Russian Propaganda, Study Finds

Key Points

• Open-weight LLMs like Nvidia's Nemotron and Alibaba's Qwen outperformed many proprietary models in resisting Russian propaganda.
• OpenAI's GPT-5.4 achieved the highest mean score of 88.9, with 54% of responses marked "Exemplary."
• Claude 3.5 Haiku (2024) scored 73.1, placing it in the bottom third of 2026 models on the benchmark.
• Google's Gemini 2.5 Pro scored 82 but showed vulnerability to malicious prompts; Gemini 3.5 Flash fell to 73.
• Performance drops sharply when models are tested in Russian, affecting Gemini 3.5 Flash, Moonshot's Kimi K2, and StepFun's Step 3.5 Flash.
• Research by Gregory Asmolov warns that Russia is attempting to steer AI outputs through BRICS collaborations.
• The benchmark underscores the need for stronger multilingual defenses against state‑sponsored disinformation.