India has positioned itself as one of the world's serious AI powers, and by several measures the case is credible. Its Digital Public Infrastructure — the stack of open, interoperable government-built systems covering identity, payments, and data exchange — is genuinely without peer at national scale. The country produces more STEM graduates annually than almost any nation on earth. Its startup ecosystem has generated AI applications across agriculture, healthcare, finance, and public services. The ambition is not theater.

The problem is foundational, and it starts before the first model is trained. Every AI system is only as representative as the data it learns from. When the people who will be affected by a system are absent from that data — because they lack digital footprints, because their languages are underrepresented in training corpora, because their labor was informal and therefore uncounted, because their healthcare was delivered in a rural clinic that ran on paper — the resulting model does not see them clearly. It sees a ghost, or nothing at all.

India's data landscape reflects its social one. The country's extraordinary linguistic diversity — hundreds of languages and thousands of dialects, many with no significant digital text corpus — means that large language models trained predominantly on English and Hindi data will perform worse, sometimes dramatically so, for speakers of Tamil, Telugu, Odia, Maithili, Santali, or Bodo. That is not a theoretical concern about fairness. It is a measurable, documented performance differential that translates directly into worse outcomes for the users those models are supposed to serve.

The same structural absence shows up in health data, financial data, and labor data. India's informal economy employs the majority of its workforce. Those workers generate almost no structured data trail. AI systems trained to assess creditworthiness, predict health outcomes, or optimize supply chains will learn the patterns of the formal economy and extrapolate badly — or not at all — to the hundreds of millions of people operating outside it. A model that cannot see informal labor does not conclude that informal labor doesn't exist. It concludes that those people don't exist, which is worse.

Gender gaps compound the problem. Women in India participate in the formal digital economy at significantly lower rates than men, a disparity shaped by device ownership, connectivity access, social norms, and safety concerns around digital identity. AI systems trained on behavioral data from digital platforms will therefore encode male-dominant usage patterns as the default. When those systems are then deployed in domains like financial services, healthcare triage, or government benefit delivery, the default becomes a bias with real administrative consequences.

None of this is unique to India. The global AI industry has a well-documented and largely unresolved problem with training data that overrepresents wealthy, English-speaking, formally-employed, urban populations. What makes India's case particularly high-stakes is the scale and the stated ambition. India is not building AI for a small market. It is building AI intended to govern, serve, and shape the lives of 1.4 billion people. The gap between that intention and the current data reality is not a footnote. It is the central engineering and ethical challenge.

There are genuine efforts underway to address it. India's AI Mission has articulated goals around inclusive data collection, and several academic and civil society initiatives are working to build language datasets for underrepresented Indian languages. The question is whether those efforts are scaled and resourced proportionally to the speed at which AI deployment is accelerating. Building a healthcare AI that serves a tribal community in Jharkhand is a harder, slower, less commercially attractive project than building one that serves urban professionals in Bengaluru. The market will not solve that mismatch on its own.

The framing that India must "build AI with all" is not a slogan about inclusion as a moral virtue, though it is that too. It is a technical argument: systems trained on incomplete data make worse decisions, and when those decisions touch public health, credit, employment, and civil administration, worse decisions have consequences that compound across generations. India has a window, right now, to build the data infrastructure that makes its AI actually work for all of its people. That window will not stay open indefinitely. The models being trained today are learning from the data that exists today, and what they learn will be very hard to unlearn.