Why Kashmiri needs its own AI

Ask an AI assistant a question in English and it answers like a well-read friend. Ask it in Kashmiri and, too often, it answers like a tourist with a phrasebook — if it answers at all.

That is not because Kashmiri is hard. It is because of how these machines learn.

A model is only as good as what it read

Today's AI language tools learn by reading enormous amounts of text and finding patterns in it. The Class 8 book puts the idea simply: "The model predicts likely words. It doesn't look anything up or compare with the world." A language with billions of web pages, books and translations gives a model a great deal to learn from. A language with little digitised text gives it almost nothing.

Kashmiri had 6.8 million speakers in India's 2011 Census (Census of India 2011). But speakers are not the same as text. Much of Kashmiri lives in speech, in poetry recited rather than typed, in two scripts — Perso-Arabic and Devanagari — and in spellings that vary from one writer to the next. For a machine, that is a language it has barely met.

It is getting better — slowly

There has been real progress. AI4Bharat's IndicTrans2, from IIT Madras, is an open translation model that covers all 22 languages in India's Eighth Schedule, including Kashmiri in both Perso-Arabic and Devanagari scripts. (IndicTrans2) That matters: open models can be studied, tested and improved by anyone, including researchers in Kashmir.

But coverage is not quality. The hard question is not does a tool accept Kashmiri? but does it get it right — and who checked?

Why it matters

When AI doesn't know a language, three things happen:

1. People are left out. A grandmother who speaks only Kashmiri can't use the tools her grandchildren use for health information, a form, or a letter.
2. Mistakes go unnoticed. A machine translation that sounds fluent but changes the meaning is worse than none, because nobody who can't read Kashmiri can tell. The Class 8 lesson on hallucination warns: "It's worse on local questions — little is written about our villages and shrines."
3. The language shrinks. If every useful tool works in English, Urdu or Hindi, children learn that Kashmiri is for home and not for the future.

What it takes to fix

Three things, and none of them can be bought from abroad:

• A corpus. Sentences in Kashmiri, lined up with their English meaning, pair by pair — the raw material any translation model learns from.
• A benchmark. A fixed set of test sentences, so anyone can see which models work for Kashmiri and which do not, instead of trusting a company's word.
• Native speakers as judges. A score from a machine is not enough. People who speak Kashmiri have to rate the translations themselves.

That is the work of Kashmir AI Research, led by AmanahVale's founder. Its open dataset currently holds 124,102 Kashmiri–English sentence pairs, aligned and filtered, and its site catalogues 115 Kashmiri language resources from 2003 to 2026. (Kashmir AI Research)

How a classroom can help

This is where schools come in, and where AI education and language preservation meet. In class, students can write short Kashmiri sentences with their English meanings, and check translations a model has made — on paper or in a class sheet. Their teacher reviews the work and sends it in for the class. A parent agrees first, because the sentences may become part of an open dataset; students never make an account on the research site, nothing they send names them, and the school and class are credited — never a child by name. (How classes take part)

Children who do this learn something no textbook can give them: that AI is made from data, that data is made by people, and that their language can be in it — if they put it there.