Ask an AI chatbot about the war in Ukraine, and the answer may depend less on the facts than on what language you typed.
Researchers put GPT, Claude and Gemini through the same test: twenty statements about the war in Ukraine, translated into 112 languages, with each model asked to evaluate each one. The result was a dataset of 67,200 responses. Sorted by language, the balance between Russia-leaning and Ukraine-leaning answers shifted noticeably. Group those languages by the countries that speak them officially, and the split tracks familiar fault lines: more Russia-leaning answers show up where public opinion of Russia is warmer, UN votes back Ukraine less often, and aid to Ukraine is thinner.
This isn't one rogue chatbot with a bad prompt. The pattern held across all three models and survived when researchers removed individual statement pairs to check it wasn't a fluke. That consistency points to something baked into the training data itself, not a quirk of one company's fine-tuning.
Search engines and social platforms have faced the same accusation for years: that they reflect the geopolitics of the language you use rather than some neutral truth. Chatbots just inherited the problem while sounding more authoritative about it.