A Tamil-speaking shopkeeper in Madurai asks an AI assistant, in Tamil, whether to extend credit to a new customer who has come twice. The AI, internally, translates the question into English, reasons about the question in English, and translates the answer back into Tamil. The answer it returns is grammatically correct Tamil. It is also, in the way it weights factors, distinctly un-Tamil. It privileges the formal-transactional framing over the relational one. It under-weights the question of whether the customer was vouched for by someone known. The shopkeeper, who would have asked a senior shopkeeper a question that started with "who introduced him," gets back a reply that starts with "what is his stated employment."
The translation was fine. The reasoning was wrong, because the reasoning was English reasoning wearing Tamil clothes.
Languages encode reasoning, not just vocabulary
There is a quiet assumption in the data-first approach to multilingual AI: that languages are interchangeable surface representations of a single underlying logic. Translate the question into a privileged internal representation, reason there, translate the answer back. This approach treats translation as the hard problem and reasoning as language-neutral.
This is not how languages actually work. Tamil reasoning, when you watch a senior Tamil professional do it out loud, sequences claims and qualifications differently from English. The verb position carries different weight. The honorifics carry information about the relational frame of the conversation that does not survive translation. The Tamil legal commentary tradition argues by a different rhetorical pattern than the English common-law one.
The same is true of Bengali, of Marathi, of Malayalam, of every Indic language. The Sanskrit philosophical tradition has technical reasoning structures, purvapaksha and uttarapaksha, the disciplined presentation of the opposing view before one's own, that have no exact equivalent in English academic argument and that produce different shapes of conclusion when used seriously.
This is not romanticism. It is observable in professional practice. A senior Telugu-medium teacher, asked a question, structures her answer differently from her English-medium colleague at the next school. A senior Marathi-medium journalist constructs a story with a different sense of where the load-bearing sentence sits.
What single-logic AI loses
When an AI does its internal reasoning in English and then translates, it loses three things.
It loses the relational metadata that is grammatically encoded in many Indic languages, the honorifics, the formality registers, the inclusive-exclusive distinctions that Indo-Aryan and Dravidian languages handle differently from each other and from English. This metadata is information about the conversation, not decoration on it.
It loses the rhetorical structure that signals what kind of answer is appropriate. A Tamil question phrased a certain way is a request for relational guidance, not for a procedural answer. An English-trained model that translates word-for-word will return a procedural answer to a relational question. The user, charitably, will think the model misunderstood. They are right.
It loses the wisdom embedded in language-specific professional traditions. The Kannada agricultural extension tradition has accumulated decades of vocabulary for talking about soil, water, and crop interaction in ways that do not map cleanly to English. The Bengali legal tradition has its own argumentative patterns in the lower courts. A model that flattens these into English first and translates out is throwing away the most valuable part of the corpus.
What multi-logical AI looks like
A multi-logical model does its reasoning in the user's language, not just its production. The internal chain of thought, when the user speaks Tamil, is in Tamil. The mid-reasoning concepts come from the Tamil professional corpus, not from the English one. The English corpus is available as reference, not as the privileged internal substrate.
Building this requires three things. First, serious training data in each language, curated from professional practice, not just translated from English. Second, evaluation benchmarks that test for reasoning quality in the language, not just translation fidelity. A model that gives a fluent Tamil answer that is logically structured like an English answer should fail the benchmark, not pass it. Third, a willingness to accept that the same question, asked in two languages, may legitimately receive differently structured answers, because the questions, in their full context, are not the same question.
Code-mixing as a feature
Indian professionals do not actually speak in single languages. A doctor in Hyderabad runs her consult in a mix of Telugu, English, and Hindi, depending on what the patient brought in, what the medical concept is, and what register she wants to land in. A magistrate in Bhopal switches between Hindi for procedural matters and English for statutory citation. A college teacher in Coimbatore code-mixes Tamil and English as a deliberate pedagogical tool.
A multi-logical AI has to handle this not as noise but as signal. The choice of language for each clause carries information. A model that normalizes the code-mix into pure Hindi or pure English before reasoning is again throwing away the most informative part of the input.
The work
If you are building Indic AI, stop measuring yourself only on translation BLEU. Start measuring whether the model's reasoning, in language, mirrors how senior professionals in that language actually reason. Recruit those professionals into your evaluation panel. Pay them well.
If you are a senior practitioner in a non-English Indian professional tradition, a doctor, a lawyer, a teacher, an administrator, your reasoning patterns are training data of the most valuable kind. The next decade of useful Indian AI will be built on whether enough of you are willing to make those patterns legible.
Multilingual is the easy goal. Multi-logical is the worthwhile one.
Join the conversation
This essay is part of an ongoing community. If it resonated, the next step is to be in the room.
Join Bharath.club → Read more essays