E-Paper
AI sovereignty as strategic infrastructure
This article is authored by Aalok Thakkar, assistant professor, Computer Science, Ashoka University.
There is a peculiar consensus in Indian technology policy today: across ministries, industry bodies, and startup ecosystems, nearly everyone agrees that India must achieve AI sovereignty. The term appears in budget speeches, mission documents, and the IndiaAI summit. Yet ask ten people what it means, and you will receive ten different answers. The silence that follows is the problem.
We do not trust bridges because their engineers are intelligent. We trust them because someone wrote down exactly what the bridge must do: How much weight, in what weather, over how many years. These are measurable, testable, and auditable specifications. AI systems deployed in public life need the same treatment, not because they are likely to collapse, but because without a prior written account of what a system is supposed to do, there is no honest way to know whether it is doing it.
Consider what most people experience when they encounter government AI. A farmer in Vidarbha opens Bharat-VISTAAR and receives a pesticide recommendation, or a first-generation beneficiary consults BHASHINI to understand a welfare document in her language. These are everyday interactions the systems were built for, and they are where the absence of specification shows most clearly.
In case of BHASHINI, the platform brings natural language access to government services across dozens of Indian languages, which matters enormously in a country where English remains the de facto language of the state. But here is what happens in practice: ask the tool to translate "aapko yeh labh nahin milega" and the output can vary from "you will not receive this benefit" to "this benefit is not available to you." To a fluent reader, the difference is meaningful. To someone navigating the welfare system for the first time, in a language that is not her strongest, it may determine whether she files an appeal or simply walks away. No publicly available specification governs which rendering is correct, what accuracy threshold triggers human review, or who bears responsibility when the translation leads to a wrong outcome. The system works, in the sense that it produces output, but it cannot be audited, because what it was supposed to produce was never written down.
These questions do not require technical expertise to ask. They require something older and arguably harder: the ability to define what good looks like before you deploy, and to hold the system to that definition afterward. This is, at its core, a humanities problem, and treating it as an engineering one is where AI governance most often goes wrong.
A lawyer drafting a contract, a historian reconstructing what a term meant to a legislature, a sociologist mapping how a policy lands differently across communities are all engaged in the same underlying discipline. They make implicit assumptions explicit. They test whether definitions hold up under pressure. They ask who bears the cost when language turns out to be ambiguous. Specification work is exactly this. It asks: What should this system do for an ordinary user, in ordinary conditions, on an ordinary day? These are not questions a model can answer for itself. They require human judgment, institutional authority, and a written record.
The stakeholders in this process are not primarily engineers. They are the district officials who will field complaints when a system misfires, the civil society organisations that understand how communities use government services, the lawyers who will eventually be asked to assign liability, and the citizens whose daily lives are shaped by systems they never chose and cannot easily question. Specification is the moment when all of them should be in the room: Before deployment, not after.
India's push for indigenous data and AI infrastructure is right, but incomplete. A foundation model trained on Indian-language datasets and hosted on domestic compute is a genuine achievement. But if no one has written down what that model may and may not do — in language precise enough to be tested — then sovereignty exists only at the infrastructure layer. The governance layer, where it matters to people, remains empty. Localising AI without specifying it merely relocates opacity. The servers are Indian. The accountability is not.
What India needs now is not more compute. It needs institutions capable of doing specification work at scale: Writing formal behavioural requirements for AI in public services, evaluating whether deployed systems meet those requirements, and revisiting both as conditions change. This is closer to what a standards body does for civil infrastructure, or what an audit function does for public finance, than what a technology ministry does. The engineering talent exists. What does not yet exist, at sufficient scale, is the interpretive capacity: the historians, lawyers, ethicists, linguists, and social scientists who know how to turn broad social intentions into stable, testable definitions, and who understand what is lost when that work is skipped.
Trust in public AI will be built the same way trust in public infrastructure has always been built through demonstrated accountability, not declared intent. That begins with writing things down.
(The views expressed are personal)
This article is authored by Aalok Thakkar, assistant professor, Computer Science, Ashoka University.
