Bridging the Epistemic Gap: Why India Needs Ethical AI Audits

Without measuring contextual failure, India’s AI governance risks mistaking computational sophistication for genuine public utility.

India’s AI governance framework has a measurement problem. The India AI Mission has onboarded more than 38,000 GPUs, allocated Rs 10,300 crore for AI infrastructure, and built AIKosh into a national dataset repository holding 7,541 datasets across 20 sectors. MeitY’s India AI Governance Guidelines, released on 5 November 2025, commit the government to fairness, equity, and human-centric design as core principles. What neither the infrastructure nor the guidelines measure is whether AI systems actually work for the populations they are meant to serve.

Accuracy sans Adequacy

India’s audit regime evaluates whether AI models are technically accurate. It does not evaluate whether they are informationally adequate for the users who receive their outputs. The two are not the same, and in India’s context, the gap between them is wide.

When a model trains predominantly on Global North datasets, it becomes accurate relative to those conditions. Deployed in India, the same model produces outputs that are internally coherent but contextually unusable. The system does not hallucinate. It answers the wrong question with authority.

NITI Aayog’s October 2025 report estimates that 490 million informal workers, who contribute nearly half of India’s GDP, remain outside formal systems of protection and opportunity. These are exactly the users most likely to receive AI-generated advice never designed for their conditions. A credit-scoring model trained on formal income histories will systematically misread the financial profile of a piece-rate worker in Tiruppur. An agricultural advisory system calibrated to temperate, high-rainfall farming data will produce technically correct but locally useless guidance for a smallholder farmer on the Deccan Plateau. In both cases, standard technical audits will rate the system as high-performing. In both cases, the user is worse off for having received the advice.

The problem extends beyond agriculture and credit. Health AI deployed at primary care centres in Jharkhand or Odisha encounters patients whose symptom profiles, dietary patterns, and disease burdens differ substantially from the populations on which most diagnostic models were trained. Welfare scheme delivery systems that use AI to screen eligibility routinely encounter documentation gaps and livelihood arrangements that formal datasets do not represent. In each of these domains, technical accuracy is the wrong metric. What matters is whether the output serves the decision the user actually needs to make.

Miranda Fricker’s concept of epistemic injustice describes this precisely. When a system fails to transmit knowledge because of whose knowledge it encodes, it actively diminishes the user’s capacity to make accurate decisions. This is not a secondary harm. It is a primary policy failure that India’s current audit vocabulary cannot name, let alone prevent.

Falling Short

MeitY’s November 2025 Guidelines rest on seven sutras. The fourth is Fairness and Equity. The second is People First. Both are stated as binding principles. Neither is operationalised.

The proposed institutional architecture, an AI Governance Group, a Technology and Policy Expert Committee, and an AI Safety Institute, focuses on risk classification, incident reporting, and technical testing. None of these bodies, as currently mandated, is designed to evaluate whether a model’s outputs serve the informational needs of the populations they reach. The Guidelines acknowledge India-specific risks, including linguistic exclusion and harms to vulnerable groups. They do not establish a standard which can measure those risks.

This is not a drafting oversight. It reflects a deeper assumption embedded in the architecture: that getting the technical infrastructure right will produce equitable outcomes downstream. That assumption does not hold in contexts where the data generating the infrastructure is itself unrepresentative. Fairness cannot be a sutra if the measurement standard to assess it is absent.

The RBI’s August 2025 FREE-AI Committee report moves closer. It recommends bias audits, explainability disclosures, and model risk management for financial sector AI, backed by board-level accountability and third-party audit requirements. It is also sector-specific. India deploys AI in agriculture, health, welfare scheme delivery, and credit access, where consequences of contextual failure are at least as severe, and where equivalent audit requirements do not yet exist.

AIKosh’s 7,541 datasets are a genuine policy achievement. Twelve startups have been selected to build indigenous large language models, including Sarvam AI’s 120-billion-parameter model and Soket AI’s multilingual foundation model, both designed to reflect India’s linguistic diversity. These are supply-side investments.

They do not address the demand-side question: whether the data being collected reflects the informational realities of users who are informal, rural, multilingual, and low-income. High-quality representative data for these populations is expensive to collect and curate. It is also a public good, meaning no private actor has sufficient incentive to produce it at the required scale. Market-driven AI development gravitates toward Western-centric datasets because they are cheaper and more abundant. This is a predictable consequence of data economics, not a coordination problem that goodwill or additional GPU capacity resolves.

The social costs of resulting bias accumulate in populations least able to absorb them: crop forecast errors for farmers without an insurance buffer, credit rejections for informal workers without a formal appeal mechanism, and diagnostic errors in health AI deployed at primary care centres where second opinions are unavailable. These costs are not currently measured, attributed, or reported anywhere in India’s AI governance architecture. Without a mandated mechanism to track and attribute contextual failure, the costs remain invisible to the institutions deploying the systems. An AI vendor whose model scores well on global benchmarks has no regulatory obligation to measure how it performs for a first-generation smartphone user receiving agricultural advice in Gondi or Tulu. The absence of that obligation is a policy choice with distributional consequence

Operational Changes

Amartya Sen’s Capability Approach offers the right measurement standard. Applied to AI governance, this shifts the audit question from “Is the model accurate?” to “Does this output expand what this user can know, decide, and act on?”

First, MeitY should mandate Ethical Impact Assessments for AI systems deployed in public service delivery, evaluating contextual adequacy alongside technical accuracy using benchmarks developed with the populations the systems serve. The proposed AI Safety Institute is the right body to house this function, but its mandate requires explicit expansion to include it.

Second, AI audit boards must include social scientists, development economists, and domain specialists alongside engineers. Evaluating epistemic adequacy is not a technical judgment. Interdisciplinary audit composition is the mechanism that the Guidelines currently lack.

Third, AIKosh should publish a public audit of what its 7,541 datasets represent across informal economic conditions, scheduled language speakers, and smallholder agricultural contexts, creating a procurement signal for targeted curation aligned with the NITI Aayog inclusion agenda.

Fourth, IndiaAI Foundation Model selection criteria should require contextual performance benchmarks alongside global accuracy metrics. Indigenous models that perform well on international leaderboards but poorly for scheduled language speakers in rural public health contexts have not met the mission’s own stated goals.

India’s AI governance framework has the scale and institutional architecture to do this well. What it currently measures, technical accuracy and infrastructure scale, is necessary but not sufficient. Epistemic adequacy is the standard that India’s AI for All vision requires. The analytical tools to operationalise it exist in economics, philosophy, and development policy. Adding them to the audit mandate is a governance choice, and one India is positioned to make.

(The author is an independent public policy researcher. Views personal.)