Why AI Keeps Replying in the Wrong Alphabet, Explained

A team of researchers from several European universities wanted to answer a simple question: when you ask an AI chatbot something in Hindi, Arabic, or Russian, how does it decide which alphabet to reply in? To find out, they opened up the model and watched what happened at each stage of its thinking. AI models process text in many steps, called layers — think of them like stations on an assembly line. At each station, the researchers checked what the model seemed to be planning.

What they found was surprising and oddly human. The model recognises your alphabet almost immediately — and it understands early on which alphabet you asked it to use. But the actual decision to write in that alphabet happens only at the very end. In the middle, through most of the assembly line, the model quietly defaults to Latin letters, the plain ABCs. So the choice of script, as researchers call a writing system, is a last-second call, not an early one.

Why should you care? More than two billion people write in Arabic, Devanagari, Cyrillic, Chinese or Korean. If AI treats Latin letters as the default setting and only corrects itself at the final moment, that is fragile. It explains real irritations: chatbots that slip into English letters mid-sentence, translation tools that mangle names, and voice assistants that answer a Tamil question in the Roman alphabet. The researchers also found that smaller models follow script instructions worse, which means the free or on-device AI you use is most likely to stumble.

The honest catch: this study explains the problem rather than fixing it. It looked at a handful of models and internal behaviour, not at a ready-made solution. But it does suggest a path — deeper, better-built models handle non-Latin scripts more reliably. In the meantime, if your AI answers in the wrong alphabet, it is not you. It is the last layer.