Imagine a woman in a government hospital in Bhopal, or Chennai, or Patna, trying to access a medical scheme she is entitled to. The online medical portal and the information about the scheme both exist, but the interface assumes a formal, stylized version of Hindi she does not speak, or Tamil in a register that belongs to textbooks rather than the street she lives on. The system is technically functional, but for her, it is a wall.
This is not a story about access to technology. It is a story about whether technology, once accessed, understands you at all.
Language, always a significant feature of AI, becomes crucial the moment it moves into healthcare, education, and public services. At that point, it helps to accelerate communication between two speakers of different languages. And yet, like the woman in Bhopal (or Chennai, or Patna), that interface does not really speak their language.
This happens because AI is trained only on one slice of society’s communications – from textbooks to television, all of it written, not verbal. A model trained on dictionaries can translate the words, but a model trained on lived language and unscripted human experience understands what the words actually signify — useful when translating metaphors, for example.
At one level then, this is about linguistic irregularities. But the implications are far wider — it becomes the means by which people are either recognized by the state or invisible to it. That gap, between words and meaning, between spoken and written language, is where millions of people fall through. It is the predictable outcome of building at speed, at scale, with the data that already exists in large quantities, which is to say the data produced by populations that were already online, in languages that were already dominant. That dominant language is of course, English. Any language model adapted from English carries its assumptions into every language database, its grammatical logic, its ways of organizing meaning, its blind spots.
In India, the scale of the problem is especially enormous. With 22 official languages, and hundreds of dialects, linguistic exclusion is not a liminal problem; it is the default condition of most digital systems. The government’s Bhashini program, the National Language Translation Mission launched in 2022, is an attempt to change that. The program builds open datasets and translation tools that work directly with Indian languages rather than routing everything through English first.
Similarly, Sarvam AI, a Bengaluru-based company founded in 2023, AI4Bharat (built by IIT Chennai’s Research Lab) and BharatGen are building large language models trained from the ground up on Indian data, in Indian languages.
Karya pays rural Indian workers, many of them women in communities with limited economic opportunity, to generate and annotate language data in over fifty Indian languages. Workers own their data and receive royalties from its use. The datasets they build are the raw material from which models that understand their languages will eventually be trained.
What Sarvam, Karya and Bhashini are building is not just a technical correction. It is an argument about whose starting point counts, and at what scale that argument matters. India has more people living outside English and standard Hindi than most countries have people. If the systems being built now the technology doesdo not account for that, they it will not fail quietly; they it will fail at population scale, in hospitals and classrooms and government offices, for decades.
The same solutions are being looked at elsewhere, in different countries and different languages.
In Latin America, LatamGPT, led by Chile’s National Centre for Artificial Intelligence, with support from governments, universities, and civil organizations across 15 countries, is building a model trained on Latin American data and contexts. The goal is to build systems that understand how language is spoken across a region where Spanish varies sharply by country and class, and where millions speak Indigenous languages, like Guaraní in Paraguay, Nahuatl in Mexico, Mapudungun among Mapuche communities in Chile and Argentina.
In Africa, Masakhane, whose name means “we build together” in isiZulu, is a grassroots research community working to build open, community-owned datasets for African languages, driven by the principle that Africans should own the data that represents them.
Much of the world’s linguistic richness is not archived neatly online. It exists in oral histories, street theatre, voice notes, road signs, and the particular way a sentence gets finished in a specific neighborhood of a specific city. Turning that into training data is slow and complicated, raising questions of consent and ownership that scraping the public internet does not. Models such as Masakhane and Karya treat those questions as central rather than incidental, which is why their work is slower and messier than the dominant model, and more accountable to the people whose language it represents.
When machines learn to understand how people actually speak, they do not just talk differently; they also listen differently. A healthcare system (such as the Microsoft-powered ASHABot) that understands colloquial Marathi can hear what a patient is actually describing. An education platform that works in Bhojpuri or Gondi can reach children in the language they think in. A government portal that speaks the way people speak can be used by the woman in the office in Bhopal, who is entitled to what it offers and should not have to translate herself to claim it.
The question of which languages machines learn to understand is also the question of which people the future is being built for. That is not a technical question.
Leave a Reply