In India, Language Is AI’s Next Infrastructure Challenge
At the India AI Impact Summit, speakers said mastering India’s linguistic diversity is essential if AI is to become as universal as digital payments.
Topics
News
- Altman Sees Superintelligence Arriving in a Couple of Years
- Inside the Debate Over AI’s Ownership Model
- In India, Language Is AI’s Next Infrastructure Challenge
- Ambani Vows $110 Billion to Build AI Infrastructure in India
- India, EU Test Digital Identity Interoperability as Twin Transition Gains Shape
- India Ranks No. 2 in AI Usage—But Who Is Being Left Behind?
When Dario Amodei, CEO of Anthropic, speaks about artificial intelligence, he often reaches for sweeping metaphors: “a country of geniuses in a data center,” “exponential curves bending ever upward,” and growth rates that seem implausible.
But in a fireside conversation with Nandan Nilekani, co-founder and chairman of Infosys Ltd, on Thursday at the India AI Impact Summit moderated by Rahul Matthan of Trilegal, the discussion turned to something more grounded and arguably more consequential: language.
In India, AI’s future may hinge less on benchmark scores and model size than on whether a farmer can speak to a machine in her own dialect and be understood.
Matthan’s prompt was simple. Anthropic had launched Sonnet 4.6 with improvements across 10 Indic languages. Why does language matter so much?
Amodei replied, “Language models are multilingual, but better in the languages they’re trained on most. India has a long tail of regional languages. If AI only works well in English, we exclude millions.”
The point was not merely technical. It was philosophical.
India is home to hundreds of languages and dialects. English, while influential, is far from universal. An AI system that performs brilliantly in English but falters in Marathi, Tamil, or Assamese does not merely underperform; it leaves entire populations behind.
Anthropic, he explained, is collaborating with Indian partners to improve data coverage for this long tail of Indic languages. Sonnet 4.6 represents progress, but not completion.
“We’re not all the way there yet,” Amodei admitted. “We want performance in regional languages to match English. This is fundamentally about access and inclusion.”
In a country where digital public infrastructure has already connected over a billion people through identity, payments, and direct benefit transfers, the next inclusion frontier may well be linguistic.
From Foundation Models to Foundation Societies
The language discussion was not an isolated technical aside. It was part of a larger question: how do you move from building powerful foundation models to building foundation societies?
Amodei acknowledged a critical distinction. AI capabilities are advancing exponentially. Diffusion, however, is slower. Even if the technology froze at today’s level, he argued, its economic impact could still multiply, if only adoption barriers were addressed.
Enter Nilekani, who has spent over a decade solving precisely that problem.
As the architect of India’s digital public infrastructure, spanning Aadhaar, UPI, and large-scale public platforms, Nilekani has grappled with a question few technologists face: how do you deploy technology at population scale?
“Diffusion is both an art and a science,” he said. “It involves institutions, policy, negotiations, trust-building, guardrails. Technology is only one piece of the puzzle.”
In his telling, AI’s challenge is not invention but implementation. Not intelligence, but integration.
100 Diffusion Pathways by 2030
If AI is to benefit billions, Nilekani believes it needs structured playbooks, what he calls “diffusion pathways.”
Building on India’s DPI experience, he described a new global initiative: 100 Diffusion Pathways by 2030.
The concept is straightforward but ambitious. A diffusion pathway is a replicable model, a toolbox, for deploying AI safely and effectively at scale. It includes technical architecture, institutional design, data governance norms, guardrails, and stakeholder coordination.
The power lies in compounding learning.
In Maharashtra’s AgriStack (MahaVISTAAR), building and stabilizing the system took nine months.
In Ethiopia, applying those lessons, it took three months.
In animal husbandry with Amul, it took three weeks.
“Learning compounds,” Nilekani said.
What began as a national digital infrastructure experiment has now become a global movement. Through Co-Develop and allied efforts, versions of DPI operate in roughly 40 countries. A recent summit in Cape Town drew 1,200 delegates from 109 nations.
The AI diffusion effort, he explained, is similarly global. The coalition includes Anthropic, Google, the Gates Foundation, UNDP, Kenya, and others.
“If we accelerate diffusion, we accelerate real-world impact,” Nilekani said.
Why AI Needs India
The conversation circled back to a phrase Nilekani has often used: “India needs AI, and AI needs India.” The first half is obvious. AI promises productivity gains, healthcare advances, agricultural efficiency, and educational transformation.
But why does AI need India?
“Because this is where we will show it working,” Nilekani said plainly.
India offers scale, political commitment, digital infrastructure experience, and a uniquely tech-positive population. It has demonstrated that large, complex systems can be deployed across 1.4 billion people.
For global AI companies, that proof matters. It is one thing to showcase demos in Silicon Valley. It is another to improve crop yields for half a billion farmers or deliver adaptive learning to millions of schoolchildren.
India could become the proving ground where AI demonstrates tangible, inclusive transformation.
Beyond English
In the end, the most striking thread in the discussion was not about compute budgets or model sizes. It was about dignity.
Language is not merely a technical feature. It is identity, agency, and participation.
If AI speaks only English, it risks reinforcing existing hierarchies. If it speaks every language well, it becomes infrastructure, like electricity or mobile connectivity.
Amodei’s vision of exponential growth, perhaps even 20–25% in a bullish scenario, depends not just on technical breakthroughs, but on broad adoption. And broad adoption depends on inclusion.
Inclusion, in India, begins with language.
The exponential may be driven by data centers. But the future of AI in India will be decided in classrooms, clinics, and farms, spoken in dozens of tongues.
If diffusion pathways succeed, and if models truly master the long tail of languages, India may not just use AI.
It may show the world how to make it work for everyone.
