Responsible AI in the Indian Context

Much of the world’s conversation about responsible artificial intelligence has been written elsewhere, for elsewhere. The principles that circulate — fairness, transparency, accountability, privacy — are sound and worth holding onto. But principles become real only when they meet the ground, and the Indian ground is unlike any other. A billion-plus people, twenty-two official languages and hundreds more spoken, vast differences of income and access, and a social fabric of extraordinary diversity. Responsible AI in India cannot be a translation of someone else’s checklist. It has to be thought through afresh.

Fairness where difference is the norm

In many discussions of AI fairness, the goal is described as treating people equally regardless of a handful of protected characteristics. India complicates this in a productive way. Here, difference is not an edge case; it is the norm. A model trained mostly on English text, or on data from a few large cities, will quietly encode the assumption that the English-speaking, urban Indian is the default Indian — and treat everyone else as a deviation from it.

This is not hypothetical. A speech system that struggles with regional accents, a document reader that fails on a regional script, a name-matching system that mishandles the sheer variety of Indian names — each of these is a fairness problem, even though none of them looks like discrimination in the usual sense. Responsible AI in India has to start from the recognition that the country is plural, and that any system trained on a narrow slice of it will serve that slice best and everyone else worse.

Fairness, in this setting, is less about erasing differences and more about ensuring that the full range of Indian reality is represented in the data, tested for in evaluation, and cared about in deployment.

Transparency that reaches the person affected

Transparency is often framed as a matter of documentation — model cards, technical disclosures, audit trails. These are valuable, but they speak to specialists. In India, the more urgent question is transparency towards the person actually affected by a decision, who may not read English, may not have encountered the idea of an algorithm, and may have no realistic way to appeal.

When an automated system decides whether someone receives a benefit, a loan or a service, responsible practice means that person deserves to know that a system was involved, in a language and form they understand, and to have a human path of recourse. Transparency that exists only in a research paper is transparency in name only. The test is simple: could the least-resourced person touched by this system find out what happened to them and challenge it? If not, the system is not yet responsible, however well-documented it may be.

Accountability in a country of intermediaries

Indian life runs on intermediaries — the local agent, the common service centre operator, the trusted neighbour who is “good with phones”. AI reaches many people not directly but through these intermediaries. This changes where accountability has to sit.

If a citizen interacts with an AI system through an agent, then responsibility is shared across a chain: those who built the model, those who deployed it, and those who mediate it on the ground. Responsible AI in India has to account for this chain rather than pretend that every user is a direct, informed operator of the technology. It also means investing in the intermediaries themselves — because they are, in practice, the human face of the machine for millions of people.

Privacy in the language people live in

India has taken meaningful steps towards data protection, and the direction is welcome. But rights on paper protect people only when they are understood. A consent screen written in dense legal English, presented to someone who reads a different language, is consent in form and not in substance.

Responsible AI here means treating informed consent as something that must be genuinely informed — communicated in the person’s own language, in plain terms, with real choices rather than a single button that says “agree”. Privacy is not only a legal construct; it is a matter of dignity, of not having one’s life quietly turned into someone else’s dataset without one’s meaningful say.

An ethic, not a compliance exercise

It would be easy to treat responsible AI as a compliance exercise — a set of boxes to tick so that a system can be declared safe. We think this misses the point. Responsibility is a disposition, a way of building that keeps asking who might be harmed, who is being left out, and who holds the power in the room.

For India, that disposition has a particular character. It has to be humble about the country’s diversity, honest about the gaps in its data, and patient enough to bring the least-served along rather than optimising for the easiest-to-reach. This is close to what we mean by the dharma of technology: not merely doing no harm, but actively directing a powerful capability towards those it might otherwise pass by.

Responsible AI in the Indian context is still being written. Our conviction is that it should be written here, by people who understand this country’s realities, and that it should measure itself by how it treats those with the least power to object. That is a demanding standard. It is also the only one worth holding.