AI Models & Russian Propaganda: New Benchmark Reveals Vulnerabilities

A new benchmark measures how easily AI language models can be fooled by Russian propaganda. Sixty models were tested with 75 questions in three languages, covering 14 propaganda narratives. Each answer received a score from one to five, with a score of one meaning the model repeated Russian talking points. Anthropic's Claude models performed best, followed by Nvidia's Nemotron 3 and Alibaba's Qwen 3.6 Plus. Mistral's models, including the new Medium 3.5, were in the bottom third. This aligns with a separate study showing Mistral had a misinformation rate of nearly 37 percent. The models had no access to web search during testing, so the benchmark only reflects the language model's ability to identify and reject propaganda. This matters because Russian networks actively feed AI systems millions of disinformation articles, and a recent campaign used an AI to spread propaganda before a major election.