The Broken Equation of ‘Atmanirbhar’ AI

At the bottom of almost every AI chat window sits a small grey line: “AI can make mistakes. Verify important info.” In an office in Bengaluru, that line is background noise, the disclaimer nobody reads twice. Step outside urban, English-speaking India, and it stops working as a disclaimer at all.

Illustration: Pariplab Chakraborty.

For a first-generation smartphone user, or a school student handed a regional-language chatbot for homework help, text on a screen tends to carry the authority of a textbook. For an adolescent learner in a town in Bihar or rural Bengal, asking a Hindi or Bengali techno-assistant to explain the relevance of a historical date or the meaning behind a biological term is akin to enquiring from a teacher something concrete and legitimate.

There is no obvious reason for the student to assume that the machine is ‘guessing’, because nobody has told them it might just be the case. When these models hallucinate, engineers write it off as an acceptable error rate. For a young adult, not with the habit of resorting to second opinions or consultations, that ‘acceptable error’ becomes a fact absorbed as truth / fact. The gap between how that error is experienced in an office and in a classroom says less about testing than about how these models were actually built.

The language shortcut

There is no obvious reason for that student to assume the machine is guessing, because nobody has told them it might be. That’s exactly what Indic AI and homegrown models are being sold on, tools finally built to speak India’s languages and understand its people on their own terms.

Over the past two years, Indian AI has been sold with the idea of being ‘Atmanirbhar’/self-reliant. Homegrown models, we are told, will finally give the next billion non-English speakers a machine that understands them. Behind that language, most of these models were built the velocity way: by translating English text into Indian languages rather than by gathering Indian-language text on its own terms.

This shortcut exists for a reason – training a language model from scratch needs millions of dollars and trillions of words of clean text, and while the English internet has been indexed for two decades, most Indian languages have a much thinner digital footprint. Living idioms, regional literature and oral histories still mostly sit in books and memory, not on scrapable websites. Digitising them with precision is slow, expensive, unglamorous work. Machine-translating the English internet into Hindi or Tamil hardly requires those tedious efforts.

This is not a hidden practice; it is written into the papers. AI4Bharat’s own technical report on Airavata, its Hindi instruction-tuned model, describes translating English datasets such as Dolly, OpenAssistant and WikiHow into Hindi using its IndicTrans2 tool, citing the scarcity of good Hindi instruction data as the reason. Sarvam AI’s OpenHathi, launched as a landmark Hindi model, was not built on an original Indian architecture; it was a Hindi vocabulary layered onto the base weights of Meta’s Llama 2.

When Ola’s Krutrim went into public beta, doubtful users were informed that it was built by OpenAI. Krutrim’s own explanation was data leakage from an open-source dataset used during fine-tuning. It is a fine line to draw, but it confirms that a model marketed as India’s own carried OpenAI’s fingerprints inside its training data. A model built this way inherits more than vocabulary. It bears heavy traces of Anglo-American sentence rhythm, American conflict-resolution scripts, and cultural references wrapped in Devanagari, Bengali or other Indic script.

The clearest cost of that shortcut shows up in safety. Making a model refuse hate speech means running it through reinforcement learning from human feedback. Building it from scratch requires mammoth sociological groundwork that only a few Indian labs are funded enough to undertake. So teams import American templates instead: AI4Bharat’s IndicAlign work, for instance, translates Anthropic’s US-designed “helpful and harmless” dataset into Indian languages and uses Meta’s Llama models to generate “safe” refusals.

American categories have no vocabulary for caste discrimination, feudal intimidation or region-specific communal dog whistles. So, audits keep finding that these models miss implicit upper-caste bias in questions about jobs or housing even as they police Western political sensibilities with confidence. The same blind spot cuts the other way: satire and everyday idioms in Bhojpuri, Tamil or Bengali often get flagged as toxic simply because they do not match the sanitized Silicon Valley ideal.

None of this was inevitable. South Korea’s Naver built HyperCLOVA X on Korean texts it already owned through two decades of running the country’s dominant search engine and content platforms, training it on some 6,500 times more Korean-language data than ChatGPT had seen. The application understands Korean honorifics and legal norms because it was in reality trained on them, not translated toward them.

The opposite happened to Scots. Wikipedia’s Scots-language edition was, for years, expanded largely by a single American teenager who did not speak the language. By the time it came to light, his edits made up close to half the entire Scots Wikipedia, written in English grammar with Scots words dropped in between them. Since web scrapers ingest Wikipedia wholesale, that error sat inside the training data of major commercial models for years.

Indian languages, thin on native digitised text and now being filled rapidly with machine-translated syntax, are exposed to the same risk at a much larger scale. Even within India, this shortcut does not fall evenly. State-backed AI initiatives speak the language of national inclusion, but development has followed a Hindi-first path in practice, with Hindi and its northern dialects consistently first in line for compute and funding. Bengali, spoken by some 300 million people, has had to make do with a fraction of that investment. Tamil, Telugu, Kannada and Malayalam fare little better despite deep literary traditions and large economies behind them. Tribal and unscripted languages such as Santali, Gondi, Kurukh and Tulu are set aside as ‘low resource’, a label that describes industry’s priorities more than it speaks about the languages themselves.

The ecological enclosure

The cloud is not weightless. It is steel, concrete, coal and water, and in India it is increasingly other people’s coal and water. State governments are competing to hand hyperscale data centres cheap land, discounted power and dedicated water lines, in parks spreading across Navi Mumbai, Noida, Chennai and Hyderabad. The math behind that is stark. Karnataka’s IT and Biotechnology minister, Priyank Kharge, told the state assembly in March that one megawatt of data centre capacity needs about one acre of land, roughly Rs 70 crore in investment, and 25 million litres of water a year. By his own comparison, five questions to ChatGPT use about half a litre. Independent researchers studying Bengaluru’s centres have put daily consumption even higher. The power for all of this still comes from a grid where coal supplied roughly 71 percent of India’s electricity last year, according to the energy think tank Ember, well above the Asian and global averages.

The Wire has itself reported on this collision, from Mumbai’s push for new dam projects to feed data centre demand to the water-stressed districts across the country now hosting server farms. What that reporting makes clear is the shape of the trade: land, power and water leave rural and peri-urban communities, and in return those communities get server racks, not services built for them. This is the paradox underneath the sovereignty of language.

The land taken and the water drawn do not go toward crop insurance tools for the farmers giving up that land or educational software for the children in the district or documentation of tribal medicines still preserved in oral memory. It goes in the direction of translating Reddit threads and American FAQ pages into synthetic Hindi sentences, fine-tuning open-source models designed in Silicon Valley, and running the customer-service chatbots quietly replacing entry-level jobs in the same towns whose water table just dropped. The sociologists Nick Couldry and Ulises Mejias have a name for this pattern: data colonialism – the use of local land, water and public subsidy to run imported architectures that do not understand local culture, cannot see caste, and confidently mislead students relying on them.

Genuine independence would look different from what is being built now. It would mean the state funding universities, rural archives and community institutions to digitise living literature and oral history on fair terms, rather than continuing to lean on machine translation as a permanent substitute; safety standards built from the Indian constitution’s own vocabulary for caste violence and communal hate speeches, instead of a translated version of an American corporate rulebook; data centre approvals that actually require closed-loop cooling, independent water audits and dedicated renewable power, alongside fair wages and real protections for the young workers across Tier-2 and Tier-3 India who label and verify this data for a fraction of what the industry is worth.

Until then, the label “indigenous AI” describes a marketing choice, not an engineering one. What is really being built is an energy-hungry machine, burning Indian land and water, to run someone else’s worldview in Indian script.

Sohini Bhattacharya is a PhD researcher and Asijit Datta is an associate professor, SCMC, Symbiosis University, Pune.

This piece was first published on The India Cable – a premium newsletter from The Wire – and has been updated and republished here. To subscribe to The India Cable, click here.

The Wire is now on WhatsApp. Follow our channel for sharp analysis and opinions on the latest developments.