Open-source generative AI models are significantly less effective at filtering Russian disinformation than closed-source competitors. A study by the state-backed Institute of the Estonian Language tested 60 systems on their ability to identify and block Kremlin propaganda.
Mistral, Europe’s flagship AI firm, saw its most advanced model rank 47th in the report. All Mistral models scored below 40% in effectiveness against malicious propaganda narratives.
Closed-source models from companies like Anthropic performed better in the evaluation. Researchers noted that vulnerabilities are particularly high when models are prompted in the Russian language. These findings raise reliability concerns for open-source AI in applications sensitive to misinformation.