Do AI chatbots actually know the news they report?

Researchers tested six commercial AI chatbots (GPT-5, Claude, Gemini, Grok) against 2,100 real news questions from BBC coverage across six languages and regions. The systems aced multiple-choice (90%+ accuracy) but stumbled badly on free-response answers, and crashed to 19–70% accuracy when questions contained false premises. The core finding: retrieval failures, not reasoning gaps, cause most errors—and every model performs worst on Hindi, with strong English-language bias in their underlying sources.