Why Rural India and AI Don't Yet Speak the Same Language

This is not a story about one farmer and one bad chatbot. It is a story about who artificial intelligence was built for — and who it silently leaves behind.

Every large language model — ChatGPT, Gemini, Claude, and the rest — learns language by reading enormous quantities of text scraped from the internet. The logic seems sound: the internet is vast, global, democratic. Surely it reflects how the world actually speaks?1

It does not.

More than 55% of all web content is in English. Hindi, spoken by nearly 600 million people, accounts for less than 0.1% of indexed content. The visualisation below makes this point more starkly than any sentence can. And Hindi is the lucky one among Indian languages.

Bhojpuri has a rich oral tradition, folk music, cinema, and a living culture stretching across Bihar, eastern Uttar Pradesh, and the Terai belt of Nepal. In the training data of most leading AI models, it is virtually absent. The same is true for Chhattisgarhi, Gondi, Tulu, Kumaoni, Mewari, and Santhali.2

The disparity has nothing to do with how many people speak a language. It is about who historically had the infrastructure, literacy, and economic incentive to put their language on the internet. Rural Indians, overwhelmingly, have not.

Technologists sometimes point out that major AI platforms now support Hindi, Tamil, Bengali, and several other Indian languages. While this is an important step forward, it obscures a deeper issue: the Hindi AI understands is often not the Hindi that most native speakers actually speak.3

Standard Hindi — taught in schools, used in newspapers, spoken on national television — is a relatively uniform, sanitised register.The Hindi spoken in everyday life is far more varied and adaptive. It blends regional dialects, modifies grammar to reflect local speech patterns, and shifts naturally between formal and informal registers, creating a linguistic richness that standardized language often fails to reflect.

The stakes are far from theoretical. AI systems are now being used to deliver agricultural guidance, support healthcare decisions, provide legal assistance, and determine eligibility for government welfare programs. If rural users cannot interact with these systems in the language they actually speak, the consequences can be severe. A farmer may lose access to a deserved subsidy, while a patient may receive inaccurate dosage instructions or misunderstand critical medical advice.

There is another mismatch that goes beyond vocabulary: the medium itself. Most large language models are, at their core, text models. But rural India is not a text-first culture — it is a voice-first one.4 WhatsApp voice notes travel faster than typed messages through villages. Instructions are given by phone call. Knowledge is passed down orally.

Speech recognition has improved dramatically. OpenAI's Whisper model handles a broad range of languages and accents with impressive accuracy. But "impressive accuracy" for a news anchor's Hindi is not the same as accurate transcription for a Rajasthani farmer's Marwari, or a tribal woman from Jharkhand speaking Santali.5 Accents, ambient noise, code-switching mid-sentence — these are everyday realities that most commercial speech models still struggle with.

The problem is known. There are serious, committed people working on it — though rarely with the resources the scale demands.

These initiatives are promising. They are also underfunded, understaffed, and operating against the grain of an industry that moves fast and profits from markets that already have money.

There is a deeper point beneath all the statistics. Language is not simply a code for transmitting information. It is the medium in which people think, feel, argue, plan, and make sense of their lives.

Consider the idea of jugaad—finding creative and practical solutions using whatever resources are available, a way of solving problems that is common in rural India.6 There is no single English word that fully captures its meaning. Similarly, decisions about land or water in many villages are often made collectively, based on relationships, trust, and shared responsibilities that no Western legal model cleanly captures.

This is not an argument against AI. It is an argument for AI that actually works for everyone — which requires acknowledging that "everyone" currently means something far narrower than it should.

When we talk about the digital divide, we usually mean access — who has a phone, who has internet, who can afford data. But there is a second, quieter divide running alongside the first: the gap between the language AI speaks and the language most of India actually uses.

Ramkhelawan gave up on the chatbot. He went back to calling his cousin in the city when he had questions. That works for him, most of the time. But it is not what the technology promised. And for the millions without a cousin in the city, it is not enough.

AI has the potential to become one of the most powerful technologies for reducing inequality in this century. But that promise can only be fulfilled if those building AI treat linguistic inclusion not as a simple diversity goal, but as a core part of designing these systems from the very beginning.

The language gap is not a small problem at the edges of AI development. It is a large problem at the centre of it — we have just been looking away.