New benchmark tests how propaganda can steer AI model outputs

The Institute of the Estonian Language has released a benchmark for measuring how susceptible AI language models are to Russian propaganda, The Decoder reports. The benchmark tests whether models can be influenced by information operations and misleading narratives.

The work matters because language models increasingly mediate search, summaries, translation, and public-information workflows. If models repeat or normalize propaganda, they can amplify manipulation at scale.

Benchmarks like this can help researchers and policymakers evaluate resilience against influence campaigns rather than only measuring general reasoning or factual accuracy.