Bangla is among the most widely spoken languages on earth. In the world of artificial intelligence it is routinely classed as a low-resource language. Both things are true, and the reason is unglamorous: models learn by reading enormous quantities of text, and there is far less digital Bangla on the internet than there is English.
What the shortage looks like
Prothom Alo gives a small, recognisable example. One evening, busy, Tamanna asked an AI assistant in Bangla to write a letter for her daughter. The letter arrived — but many sentences felt mechanical, and in places the word choices did not fit.
That is what thin training data produces: wrong spellings, stiff constructions, translations that are technically correct and tonally wrong. The model is not refusing Bangla. It has simply read less of it, and most of what it read was translated from English rather than written by someone thinking in Bangla.
The cost nobody mentions
There is a second penalty worth knowing about, because it shows up on a bill. Models break text into tokens, and the tokenisers behind most large models were built mainly on English. Bangla script fragments into many more tokens per word than English does. The same question therefore costs more to ask in Bangla, fits less history into the model’s context, and runs slower.
So a Bangladeshi developer building in Bangla pays more per user than one building the identical product in English. That is not anyone’s policy; it is a side effect of whose text the tokeniser was trained on. It is also a real commercial disadvantage, and worth naming when people wonder why local-language products are rarer than local-language demand.
What is changing
Researchers at home and abroad are building models specifically for Bangla, and Bangladesh’s draft National Artificial Intelligence Policy (2026–2030) speaks of responsible and inclusive AI development. A policy document does not generate training data, but it does decide whether public text — government forms, court judgments, textbooks, archives — becomes openly available for anyone to train on. That single choice would do more for Bangla in AI than any number of statements of intent.
What to do in the meantime
Ask in short, clear sentences. Read what comes back and fix it yourself — always for job applications, legal documents and office letters, where a stiff sentence or a wrong word carries a cost. Do not paste an AI-written Bangla letter into anything that matters without reading it aloud once.
And the part that sounds sentimental but is mechanically true: the more a language is used, the better machines learn it. Every piece of correct Bangla written online — a blog post, a product description, a comment, a page of documentation — becomes training data for whatever comes next. The question is not only whether AI will learn Bangla. It is whether we will write enough of it, well enough, for there to be anything worth learning from.
One thing worth not doing
A common response to weak Bangla output is to write the prompt in English and ask for a Bangla answer. It usually produces smoother sentences — and that is the problem. What comes back is English thinking in Bangla clothing: the idiom is wrong, the sentence order is foreign, and the result reads like a translation because it is one. For anything that will be read by Bangla speakers who did not ask for a translation, it is better to accept a stiffer draft written from a Bangla prompt and fix it yourself.
