EU Study Finds LLMs Parrot Russian Propaganda — But Only When You Ask in Russian
A European Union study found that major language models repeat Russian state narratives about the Ukraine war 86% of the time when queried in Russian, while producing balanced responses in English and Ukrainian — exposing a severe language-dependent alignment gap.
A study commissioned by the European Union has produced one of the most concrete demonstrations yet of language-dependent bias in large language models. As @petite_michelle reported, when researchers queried leading LLMs about the Russia-Ukraine war in Russian, the models reproduced Russian state propaganda narratives 86% of the time. The same queries in English or Ukrainian produced substantially more balanced responses.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.