India’s AI opportunity is not to win the race to the most advanced model. It is to show the world how AI can be made useful to real people, solving real problems. That is what India’s own strategy says, and what framed the 2026 India AI Impact Summit. India has the ingredients: linguistic and cultural diversity, digital public infrastructure at scale, a young technical workforce, and startups working on problems that matter.
Two things stand in the way.
The first is that we are measuring impact the wrong way. In the current policy discourse, impact has come to mean adoption and diffusion—how many sectors, how many users, how many languages. These are useful numbers, but they say nothing about whether the systems being adopted actually deliver, and nothing about the new vulnerabilities they create. A system that performs adequately where institutional oversight is strong can cause serious harm where that oversight is thin. The same conditions that let India leapfrog also leave it exposed. Impact has two halves; we are counting one.
The second is dependency. India’s ecosystem of context- and domain-specific models is genuinely vibrant, but almost all of it sits on top of a handful of American labs. Dependency in itself is not the problem; it exists in many sectors. It becomes a problem when you cannot negotiate its terms. AI credits have meanwhile become a form of development assistance, as seen in agreements being signed between leading AI companies and global philanthropies.
Neither problem has a single fix. Both will need competition policy, public compute, better redress mechanisms and much else. But one piece is common to both, and it is the piece India has invested in least: independent evaluation.
To define impact fully, India must be able to see how the systems it adopts actually perform across its languages and settings, and to detect the harms adoption produces. To manage dependency, it must be able to set conditions on what it accepts, and to act when those conditions are not met. Both turn on the same capacity.
This makes evaluation strategic rather than merely technical. It is the layer at which India can assert sovereign authority even where the underlying models are not sovereign. Sovereignty in AI need not mean building the models. It can mean choosing between them on terms India has set.
Procurement is where this becomes real. The Indian state is unusual in the scale of its role as an AI adopter: among the largest purchasers, a builder of public compute and datasets, and a funder of indigenous foundation models. A procurement regime that requires independent evaluation—demonstrable performance in Indian languages, evidence of safety in the deployment context, post-deployment monitoring for high-stakes uses—turns the State’s buying power into a tool for raising market standards. The light-touch approach of the November 2025 AI Governance Guidelines makes this more urgent: rules that rely on principles and self-regulation only work alongside continuous evidence generation.
None of this is easy, and the field is far from settled. A systematic review of 445 large-language model benchmarks found only 16 percent used rigorous methods to compare model performance, and roughly half claimed to measure abstract qualities like “reasoning” or “harmlessness” without defining them. The problems compound across languages: test datasets do not exist for many low-resource languages, and translating English ones fails to capture how people speak or prompt.
Even good benchmarks are not enough, because the most consequential effects of AI are not visible at the model level. They emerge as people use systems over time. A frontline health worker who routinely defers to an AI recommendation may lose judgement she once exercised. A welfare officer who overrides an algorithmic flag may face pressure to do so less often. Beyond these interaction harms sit systemic ones—effects on labour markets, on human agency, on cognitive ability.
This is where the gap is starkest. Serious work is now under way in India on model performance, including multilingual benchmarking. On interaction and systemic harms, there is almost nothing—a global survey of generative AI evaluations found fewer than 6 percent accounted for human-AI interaction. These harms need observational, longitudinal, often qualitative methods, and domain knowledge more than technical knowledge. They are slower, costlier and harder to reduce to clean numbers, which is why they are barely funded.
Public trust is also a factor. Across much of the West, enthusiasm for AI has given way to suspicion. India is in a different position: public sentiment here remains optimistic about technology. That optimism is an asset, but it is not permanent. It will hold only if these systems prove safe as well as useful, and if failures are caught by someone other than the companies that built them. Evaluation is how that gets demonstrated.
India should also choose its niche. Frontier capability testing is already being done in the UK and the US, which have the compute and secure infrastructure for it; there, India needs partnerships for joint testing and timely sharing of results. Its comparative advantage lies in the application layer, multilingual evaluation and post-deployment monitoring—work every country in the Global South will need, which makes it exportable.
Three steps would get us started. Fund application-layer sandboxes, as Singapore does, where third-party evaluators test commercial AI applications and the state turns findings into public guidance, but built around domain communities of practice, with doctors, teachers, farmers and legal aid workers designing the rubrics. Focus the AI Centres of Excellence in health, agriculture and education as certification and assurance bodies rather than pushing them only towards commercial products; assurance is the clearer public-interest mandate.
Third, and most neglected: grow an ecosystem of independent third-party evaluators. None of this works without people to do the work, and India today has very little capacity—some at technical institutions, some in civil society, nothing near what the task requires. The field has to be actively cultivated: training, shared methods and tooling, and institutional homes where an evaluation career is possible. It should be a named budget line for governments and development finance agencies funding AI adoption, not an afterthought.
Civil society organisations belong at the centre of that ecosystem, not at its edges. They see systems in use, in context, over time—a vantage no laboratory has, and exactly what interaction and systemic harms require. Their absence from formal evaluation says more about how we have defined technical expertise than about what the work demands.
Finally, evaluation is labour. Building datasets and running safety tests exposes people to disturbing content, as content moderation does. Fair pay, psychological support and transparent contracting belong in any publicly funded evaluation programme.
India wants to be the ‘AI use capital’ of the world. It cannot be that without also becoming its evaluation capital.
Urvashi Aneja | Founder and Director, Digital Futures Lab and Co-Convenor, Global South Network for Trustworthy AI
(Views are personal)