If you import the world’s best car and put it on a road that was never built for it, the problem is not the engine. The problem is the system around it.
India is not short on the engine. Between global frontier models and a fast-growing set of homegrown ones, Indian learners already have more artificial intelligence (AI) capability than any generation before them. What decides whether that capability improves outcomes for the country’s students and its AI-disrupted workforce is not which model wins, it is what gets built between the model and the learner. Right now, that work is underspecified, and most of the conversation about AI in Indian education is still arguing about the engine.
The Real Language Problem Is Arithmetic, Not Sentiment
Most large language models still break Hindi, Tamil or Kannada down into more computational units than the same sentence in English. That’s because the tokenisers underneath them were shaped by training data that skews heavily toward English and Latin scripts.
That is not a rounding error. Token count determines cost and speed, so the same model can become slower and more expensive for the learner who is thinking and typing in an Indian language. In a country where affordability is the single-biggest barrier to who gets to use AI, that arithmetic decides who this technology reaches, long before anyone discusses pedagogy.
This is precisely the gap a new generation of homegrown language models is trying to close. Tech Mahindra’s Project Indus, an 8-billion-parameter Hindi-first model, and Sarvam AI’s open-weight Sarvam-30B and 105B models are explicit attempts to fix the economics of running AI in Indian languages, not just the vocabulary.
Any AI education company serious about affordability should be building toward a multi-model future. Routine interactions must run on whichever localised model handles that language most cheaply, while global frontier models are reserved for the harder reasoning tasks where their extra capability earns its cost. Sending every basic query to the single-most powerful available model is not rigorous. It is an expensive habit dressed up as one.
Cultural fit is the easier problem: examples that reference an Indian classroom, instead of an American one, a tone that does not feel imported.
The harder problem is narrative: whose account of a historical event, whose assumptions about what counts as a well-reasoned argument get embedded in an answer delivered with the same flat confidence as a fact. A model trained overwhelmingly on English-language, largely Western internet content does not just miss an Indian reference occasionally; it carries a specific worldview into a classroom without announcing it. That is a different order of risk from a bad example because a bad example is obviously wrong. A quietly embedded narrative is not obviously anything. It just reads as the answer. This is not a translation problem a localisation team fixes by rewriting prompts. It requires Indian educators reviewing generated content for framing, not only for factual accuracy or tone, before it reaches a learner at scale.
Trust Has a Deadline Now
Data trust in Indian education is no longer a value statement. It has a calendar attached to it. India’s Digital Personal Data Protection (DPDP) framework takes a lighter touch on moving data outside the country than Europe’s GDPR: transfers are generally allowed unless the government restricts a country or entity, rather than being blocked by default. That is more permissive than most platforms assume, and it means genuine trust depends on the platform’s own design choices, not on the law doing that work automatically.
Crucially, the formal Consent Manager Framework becomes operational on November 13, 2026. Any company still treating that as a distant compliance milestone rather than a near-term architectural deadline will be forced into rushed, expensive overhauls this year that should have been native from day one.
The Unknown Risks
A fourth risk gets almost no airtime, and it has nothing to do with any model’s quality: dependency. Building critical education infrastructure on one foreign vendor’s application programming interface (API) means access can be shaped by decisions made far outside India.
On June 12, 2026, the US commerce department issued an unprecedented directive restricting foreign access to a leading US AI lab’s frontier models. Access via public APIs was instantly tangled in export controls, disrupting platforms relying on them until the restrictions were adjusted weeks later.
Multi-model redundancy is no longer just good engineering; it is risk management.
Nobody in India made that decision, or could have stopped it. This is not a hypothetical; it happened to real, production-level products. The practical answer is not to abandon global models, which remain ahead on the hardest reasoning tasks. It is to ensure that no single vendor, foreign or Indian, ever becomes a single point of failure for a product millions of students depend on daily. Multi-model redundancy is no longer just good engineering; it is risk management.
India’s own AI ecosystem has moved fast and deserves real credit. But an honest assessment has to hold homegrown claims to the same standard as foreign ones.
Several of the strongest capability benchmarks published by new Indian models have not yet been independently verified, and India does not yet have a trusted, domestic equivalent of a Chatbot Arena or Hugging Face to arbitrate them. This is no criticism of them. It is a reminder that “built in India” and “proven to work at scale” are two different sentences. A platform putting a model in front of a student should be able to answer the second question with independent evidence, not brand loyalty.
None of this argues against using powerful global models. It argues for using them deliberately. Concretely, that means four things:
Route intelligently: send routine interactions to whichever local model handles the language most efficiently, and reserve frontier models for the reasoning tasks that need them.
Design for November 13, 2026: treat this year’s DPDP Consent Manager activation as the hard deadline for data-handling architecture.
Audit for framing: have Indian educators review generated content for structural narrative, not just for basic accuracy.
Verify blindly: refuse to take any vendor’s benchmark claims, homegrown or global, as settled until independently tested.
Go back to the car. The engine was never the hard part. The world’s AI labs are building extraordinary engines. What decides whether any of it reaches the student who has never had access to a high-quality tutor is the unglamorous work: the routing, the pedagogical review, the hard regulatory deadlines and a refusal to take anyone’s word for it. That is the work. It is not exciting to write about. It is the only part that actually matters.
(Views expressed are personal)
Anshuman Singh is the co-founder of Interviewbit And Scaler, one of among India’s largest AI native education companies
(This story appeared in Outlook magazine’s August 3 issue, 'The AI Divide', which focuses on how India's AI education ambitions are colliding with the reality of inadequate digital infrastructure, undertrained teachers and AI tools that are not built around Indian students' cultural context)
