TThe Hindi LLM Lie

Open any major foreign AI product and you will find a familiar claim somewhere in the marketing material: "now supports Hindi." Try the claim. Type a question in everyday Hindi, the kind a small business owner in Lucknow might ask their accountant, with the proper mix of formal grammar, colloquial idiom, and English loanwords. Watch what comes back. In most cases, you will get a response that is technically Hindi, the script is right, the grammar is approximately right, and yet immediately, unmistakably, the response of someone who learned the language from a textbook in another country. The register is wrong. The honorifics are wrong. The cultural references are wrong. The English residue is everywhere.

This is what most "Hindi support" actually is: an English model with a translation wrapper, sometimes a thin layer of Hindi fine-tuning, marketed with a confidence that the demo will not be probed. For an Indian user, the experience is uncanny, the model speaks Hindi the way a foreigner speaks Hindi after a one-year language course. Competent. Fluent in dictionary terms. Wrong in the ways that matter.

¶Why this is not a small problem

The reason this matters is that Hindi is not a small language. It is the working language of three to four hundred million Indians, depending on how you count. It is the primary language of education, government, and commerce in a strip across north India that contains some of the country's largest professional populations. Hindi at the depth that working professionals actually use is the cognitive layer of a sub-continent.

An AI model that produces fluent-but-foreign Hindi has a specific failure mode in professional contexts. It will hallucinate confidence on cultural references it does not understand. It will use the wrong honorifics, which feels insulting in formal contexts. It will misinterpret idioms, sometimes catastrophically. It will produce legal and medical advice in a register that no Indian lawyer or doctor would actually use, undermining the credibility of the advice even when the substance is correct. In any high-stakes setting, the failure of register is not cosmetic. It is operational.

¶Native depth versus translation depth

The distinction that matters is between native depth and translation depth. A model with translation depth has learned to map Hindi tokens to English tokens and back. It can render a sentence in either direction. It can answer simple questions. It cannot reason in Hindi. The reasoning happens in the English latent space, and the Hindi is layered on at the end. The result is a thinking machine that thinks in English and speaks in Hindi.

A model with native depth has been trained on enough high-quality Hindi text, and code-mixed Hindi text, and conversational Hindi audio transcripts, and Hindi reasoning chains, that its internal representations of common Indian concepts are anchored in Hindi as much as in English. When asked a Hindi question, it reasons in Hindi. The register, the idiom, the cultural context come naturally because they are not being translated; they are being produced from a layer that knows what they are.

The gap between these two kinds of models is enormous, and from the outside, you cannot tell which kind a given product is. The marketing always says "supports Hindi." Only sustained use reveals the difference.

¶What native Hindi data actually requires

Building a model with native depth requires data that almost no scraping pipeline currently produces. Conversational Hindi from actual professional contexts. Code-mixed Hindi-English from offices, clinics, and shops. Hindi reasoning chains where someone thinks aloud through a problem in Hindi, with all the digressions and verifications a real thinker uses. Multilingual transcripts where the conversation switches between Hindi and Tamil or Hindi and Marathi as Indians actually do in mixed-language settings.

This data has to be created, mostly, because the public Hindi internet, which dominates the easy data, is heavily Wikipedia, news, and translation. The data that captures how Hindi is actually used in working life is in private contexts, and producing it at scale is expensive, slow, and culturally delicate. It cannot be done by a Bay Area team commissioning translators. It can be done by Indian institutions and communities that have access to the linguistic textures they are trying to capture.

This is the part of the Indic AI story that gets the least public attention and is the most strategic. The team that produces, over the next five years, the highest-quality native-depth Hindi training data, and the corresponding datasets for Tamil, Telugu, Bengali, Marathi, and the other major Indian languages, will have effectively produced the substrate for a generation of native-quality Indic AI. That team will not be a single company. It will, in all likelihood, be a network of universities, professional communities, and language-specific groups, coordinated lightly, working on long timelines.

¶The code-mixing problem

A subset of the native depth problem deserves its own attention: code-mixing. Indian professional speech, especially urban speech, switches between Hindi and English fluidly, sometimes inside a single sentence. "Aaj client meeting hai, presentation ready hai kya, ya last minute mein scramble karna padega?" That sentence is grammatically Hindi with English nouns and verbs woven in. It is the way three hundred million Indians speak at work. Almost no foreign model handles it natively. Most fall back to one language or the other, missing the texture entirely.

A model that handles code-mixing natively can be used by professionals in their actual working register. A model that does not is one more tool that requires the user to translate themselves into the model's preferred mode before the conversation can start. The cost of that small translation, multiplied across millions of professionals, millions of times a day, is enormous and invisible.

¶The opportunity is the work

The Hindi LLM lie persists because there has not been enough public pressure to call it out. The marketing keeps moving faster than the verification. The labs in foreign countries continue to declare Hindi support, and the demos continue to look fine to people who do not speak Hindi well enough to probe them. The fix is not better foreign labs. The fix is Indian teams, Indian linguists, Indian users, and Indian communities producing native-depth Hindi models and the evaluation pipelines that distinguish them from the translation wrappers.

Sarasvat.ai exists, in part, to do this work. AI.Bharath.CLUB exists, in part, to provide the community whose feedback makes the work possible. The lie will not be defeated by argument. It will be defeated by the slow accumulation of better models, used by more professionals, in more contexts, until the difference between native and foreign-with-costume becomes too obvious for any marketing copy to paper over.