Authors 

Gaurav Godhwani & Ahsnat Mokarim

Abstract

AI systems are only as effective as the data on which they are built. In India, data is abundant but remains fragmented, uneven in quality, and poorly structured for reuse across systems, creating a widening gap between data availability and data usability. This essay argues that the core constraint in India’s AI strategy is not the lack of data, but the need for a coherent data infrastructure. It identifies five key structural challenges: poor data quality, limited interoperability, gaps in representation, weak governance frameworks, and consent mechanisms. It shows that current policy and platform efforts have improved access to data but have not addressed the conditions required for its effective use. In response, the essay proposes a shift towards a coordinated data ecosystem built on responsible data infrastructure, clear standards, incentive-based data sharing, domain-specific data collaboratives, and institutional data stewardship.

The Data Deficit in India’s AI Strategy 

The IndiaAI Mission, approved by the Union Cabinet in March 2024 with an outlay of INR 10,371.92 crore over five years, sets out a plan across seven pillars around India’s AI ecosystem development: computing infrastructure, data, talent, research and development, capital, algorithms, and applications. Platforms like AIKosh have onboarded over 10,000 datasets and 294 AI models across 20 sectors. These include government datasets relating to health, agriculture, geospatial mapping, mobility, education, and welfare administration. On the language front, BHASHINI supports approximately 22 Indian languages through more than 350 AI models. BharatGen extends this to all 22 scheduled languages. The IndiaAI Compute Portal has scaled to over 38,000 GPUs at subsidised rates.

Despite these achievements, India’s strategic approach has not significantly prioritised the importance of data as the foundational input for accurate AI performance. As per the Economic Survey 2025-26, India generates nearly 20 per cent of world’s data, however, much of this is raw transactional data rather than the curated and representative material that AI systems require. India also hosts only around 3 per cent of global data centre capacity, so a large share of its data is stored and processed abroad. Of the IndiaAI Mission’s INR 10,371.92 crore budget, only about 1.9 per cent is allocated to the IndiaAI Datasets Platform. Institutional and financial priorities have instead centred on compute infrastructure, model innovation, and startup ecosystems. 

Figure 1: Budget allocation across the IndiaAI Mission’s seven pillars

The data layer, particularly issues of data quality, governance, and responsible data sharing, remains critical in determining both the reliability of AI systems and the distribution of their benefits. The performance of any AI system is determined by the quality, completeness, and representativeness of its training data. A model trained on fragmented or demographically skewed data produces outputs that are unreliable, exclusionary, and difficult to correct. In India’s public sector context, where AI is being deployed in welfare delivery, agricultural advisory, health diagnostics, legal administration, and disaster management, these outputs fall not on abstract users but on specific communities, many of whom are already the most underserved by existing public systems. Strengthening India's data infrastructure should therefore be treated as a precondition for achieving India’s AI objectives. 

Structural Challenges in India’s Data Pillar

Data Quality, Fragmentation, and AI Readiness 

Government datasets across sectors are inconsistently structured, incompletely documented, and rarely organised in ways that AI systems can use for training and evaluation without substantial manual preparation. NITI Aayog attributes this to systemic issues such as (i) data entry driven by speed and enrolment targets rather than accuracy; (ii) no clear ownership of data within or across departments, so errors persist uncorrected; and (iii) almost no real-time validation at the point of capture. These shortcomings have consequences beyond administrative inefficiency. Erroneous or duplicate beneficiary records contribute to fiscal leakage, with NITI Aayog estimating that they increase welfare expenditure by 4–7 per cent annually. Inconsistent and non-standardised datasets also create policy blind spots by distorting the evidence on which governments rely, leading to poorly targeted interventions and delayed policy corrections. For citizens, inaccurate records can result in rejected claims and lengthy rectification processes, gradually eroding trust in digital public services. When these same datasets are used to develop AI systems, their deficiencies are reproduced at scale, increasing the risk that automated decisions reflect errors and biases embedded in the underlying data rather than actual social conditions.

Beyond data quality, AI systems also depend on whether datasets are AI-ready. Comparative frameworks, such as the United Kingdom’s Guidelines for AI-ready Datasets, emphasise that AI-ready datasets require high-quality metadata, interoperable data structures, clear governance arrangements, and legal and ethical safeguards throughout the data lifecycle. While initiatives such as AIKosh have expanded access to datasets, India has yet to establish a comparable framework for systematically assessing whether datasets meet these standards. The challenge, therefore, is not only to make more data available but to ensure that existing datasets are consistently prepared for effective, reliable, and accountable AI deployment.

Interoperability

Interoperability is a structural challenge across the data ecosystem. Cross-sectoral AI applications that require connecting different sectoral datasets are structurally difficult to achieve. For example, healthcare AI applications may require anonymised health records with nutrition data, or agricultural AI applications may require weather data with soil quality records. These challenges are not only technical but also organisational and legal, involving administrative compliance, incompatible legacy processes, and the absence of shared governance for interoperability initiatives. Many of the constraints are also infrastructural, since much of the existing data infrastructure was built for static record-keeping rather than the continuous, machine-readable exchange that AI development requires.

These constraints are especially acute within government data systems built in departmental siloes, with limited incentives for responsible data sharing. Departments face institutional risks in sharing data, including concerns around privacy, compliance, and commercial interests, while the benefits remain uncertain and unevenly distributed. NITI Aayog notes that weak ownership and accountability structures reduce incentives to maintain or share data across systems. Many public authorities also lack the institutional capacity, common standards, and governance arrangements needed to curate, document, and maintain datasets in forms that can be readily reused across departments. Preparing datasets for external use therefore imposes significant administrative costs without creating clear institutional benefits, which reinforces a risk-averse approach to data sharing. Studies of open government data in India have also identified institutional reluctance to release datasets, noting that concerns over control, accountability, and public scrutiny can discourage proactive data sharing. The persistence of data silos therefore reflects not only technical barriers but also the way institutional responsibilities, capacities, and incentives are currently organised. 

Representation and ExclusionWhile fragmentation and interoperability are supply-side problems concerned with the movement of existing data, the third challenge is one of representativeness. It concerns data that does not reflect the lived experiences of the communities that AI applications are meant to serve. For example, researchers from the University of London and Edinburgh found that AI models trained on ICU datasets underestimated clinical deterioration in markers of prognosis among women by up to 33 per cent, as compared to men. They also found that the models were producing worse outcomes for non-White patients. These distortions are also seen in the datasets that developers rely on in practice in India. Research on India’s healthcare AI ecosystem shows that, in the absence of high-quality public datasets, many developers depend on publicly available datasets from the United States and Europe to train and validate AI systems. While these datasets may be valuable for research, they reflect different demographic profiles, disease burdens, clinical practices, and institutional contexts, increasing the risk that AI systems perform poorly when deployed in India. 

The consequences extend beyond model performance. Limited access to high-quality Indian datasets also creates unequal opportunities within the AI ecosystem. Smaller startups, academic researchers, and public-interest innovators often have little choice but to rely on publicly available datasets, regardless of their suitability. By contrast, large technology firms and established industry actors possess the financial resources and institutional relationships needed to negotiate access to data held by major public and private hospitals. As a result, the scarcity of accessible, high-quality datasets reinforces existing market asymmetries, concentrating AI development among actors that already control valuable data resources while raising barriers for new entrants. This also weakens India’s ability to build competitive advantage from the data generated within the country. 

Consent

The challenge of consent further compounds these problems. This is particularly important where AI systems involve sensitive personal data used in sectors such as healthcare, welfare, and education, where they increasingly depend on the secondary use of information. Much of this data is originally collected for a specific administrative or service-delivery purpose but is later reused for AI development, model evaluation, or research. This exposes a limitation of conventional consent frameworks. Procedural consent, obtained through a signature or click-through agreement, may satisfy legal requirements, but it does not necessarily ensure that individuals understand how their data may be reused, who may access it, or the risks associated with downstream AI applications. Meaningful or substantive consent requires that individuals are able to understand these future uses and exercise genuine agency over them. However, this becomes increasingly difficult as datasets are linked, shared across institutions, and repurposed for uses that were neither foreseeable nor specified at the time of collection. 

These limitations become especially acute in the context of community data. For example, the Government of India recently confirmed that data collected under the Scheme for Protection and Preservation of Endangered Languages and Sanchika’s government archives were used to enrich training for AI models. However, it remains unclear whether the communities whose languages were originally collected under a preservation mandate consented to their subsequent use in AI training. While the Digital Personal Data Protection Act, 2023 establishes rights over individual personal data, it does not recognise collective linguistic or cultural interests, and communities whose knowledge contributes to AI systems consequently have little influence over how that data is subsequently governed, shared, or reused. Addressing these gaps therefore requires moving beyond consent as a procedural requirement towards governance mechanisms that ensure transparency, accountability, and meaningful participation throughout the lifecycle of AI data. 

Figure 2: Key data and governance challenges shaping India’s AI ecosystem.

These challenges show that India’s data deficit is not simply a question of insufficient data, but of whether existing datasets can support reliable, accountable, and socially useful AI systems. Investments in compute infrastructure, foundation models, and AI applications cannot fully compensate for fragmented, poorly governed, or unrepresentative data. Weak data infrastructure is therefore a fundamental constraint on AI capability. It limits model performance, reduces opportunities for independent evaluation and public scrutiny, and reinforces existing inequalities in AI outcomes in the distribution of AI’s benefits and harms. Addressing these deficits requires more than expanding the volume of available datasets. It requires governance arrangements that improve how data is collected, managed, shared, and reused across institutions. The following section therefore examines the extent to which India’s current data governance architecture is equipped to meet these challenges.

India’s Current Data Governance Architecture and Its Gaps

The Legislative and Regulatory Layer

The Digital Personal Data Protection Act 2023 is India’s first comprehensive data protection legislation. It recognises individual rights to information, correction, erasure, and grievance redressal, and places obligations on data fiduciaries around consent, purpose limitation, and data minimisation. It regulates how digital personal data relating to identifiable individuals may be collected, processed, and used by entities designated as “data fiduciaries”, and establishes rights for individuals, including consent-based processing and obligations around data security. However, the scope of the Act is limited to the governance of personal data. In other words, it does not address the management or use of non-personal, anonymised, or aggregated datasets. 

Alongside this law, India has developed several sectoral initiatives intended to enable data sharing and digital innovation. These include the Data Empowerment and Protection Architecture (DEPA), the Account Aggregator framework, and language data initiatives such as Bhashini. While these initiatives address particular domains, they do not create a general framework for governing datasets that may later be used in AI systems. Questions relating to public-interest use of shared datasets, including community language resources, ecological information, or agricultural knowledge, therefore remain only partially addressed.

An attempt to address this gap appears in the Draft National Data Governance Framework Policy, 2022, issued by the MeitY. The draft framework proposes mechanisms to facilitate the sharing of non-personal data for research and innovation. These mechanisms, however, rely largely on voluntary participation by institutions that hold data. In practice, organisations that maintain high-quality datasets often invest significant resources in generating and maintaining them and may face reputational or institutional risks if imperfect or incomplete data is released publicly. Without clear legal obligations or incentives, many institutions may therefore remain reluctant to share such datasets. The pool of data that becomes available may consequently reflect the practices of actors already comfortable with open data, while datasets held by state agencies, civil society organisations, and community-level actors remain underrepresented.

A related development is the set of principles issued by the MeitY through the AI Governance Guidelines, 2025. These guidelines emphasise transparency and accountability in the deployment of AI systems. While they refer to the use of datasets in training such systems, the primary focus of the document is the behaviour of AI systems after deployment. The guidelines do not establish concrete institutional or procedural mechanisms to govern the datasets used in training. As a result, there is limited oversight over questions of data quality, representativeness, and integrity, even though these factors directly shape whether AI systems are fair, reliable, and auditable.

Further efforts to expand access to public datasets can be seen in the Draft India Data Accessibility and Use Policy, 2022. The draft policy proposes to make datasets generated by government ministries and departments more widely accessible for research, innovation, and policy development. It suggests an institutional structure that would include an India Data Office along with departmental data management units responsible for identifying and managing high-value datasets. The proposal responds to the long-recognised problem that large volumes of administrative data exist within government but remain scattered across departments and are difficult to access in usable formats. At the same time, the draft policy does not clearly explain how this proposed framework would interact with the various data portals and sector-specific repositories that already exist. In practice, it risks adding another layer to an already fragmented data ecosystem rather than resolving the coordination challenges that currently shape access to public datasets.

In sum, India’s existing legislative and policy instruments address different aspects of the data and AI ecosystem. The Digital Personal Data Protection Act, 2023, regulates the processing of personal data. The Draft National Data Governance Framework Policy, 2022 proposes mechanisms for sharing non-personal data. The AI Governance Guidelines, 2025 articulate principles for accountability in AI deployment. However, these instruments operate largely in parallel. They neither explicitly recognise collective or community interests in data, nor establish enforceable standards relating to the governance or quality of datasets used in AI development. They provide limited assurance to organisations and communities about how shared data will be managed, reused, and governed once it moves across institutions, which may discourage sustained participation in data-sharing arrangements. As a result, the framework does not yet function as an integrated system of data and AI governance.

The Platform and Infrastructure layer

The National Data and Analytics Platform (NDAP), launched by NITI Aayog in 2022, brings together datasets from multiple ministries and sectors and provides tools that allow users to search, visualise, and combine them across government sources. The platform was conceived as a way to make public data easier to access and analyse, particularly where datasets had previously been scattered across different departmental portals.

Alongside this, the Open Government Data Platform India, operated by the Ministry of Electronics and Information Technology (MeitY), functions as a central portal for publishing shareable government datasets in machine-readable formats. Developed and maintained by the National Informatics Centre (NIC), the platform allows ministries and departments to upload datasets that can be searched, downloaded, and reused by researchers, businesses, and civil society. Several government agencies also maintain specialised sectoral data portals. For instance, Bhuvan provides geospatial datasets, while the India Water Resources Information System hosts hydrological and water resource data.

A different approach to data sharing appears in the Data Empowerment and Protection Architecture introduced in 2020. DEPA establishes a consent-based mechanism through which individuals can authorise the transfer of their personal data from institutions that hold it, such as banks or tax platforms, to third parties providing services like financial management or credit assessment. These transfers occur through regulated intermediaries known as consent managers. The model has been implemented most prominently in the financial sector through the Account Aggregator ecosystem under the supervision of regulators, including the Reserve Bank of India, Securities and Exchange Board of India, Insurance Regulatory and Development Authority of India and Pension Fund Regulatory and Development Authority. In effect, DEPA places individuals at the centre of decisions about how their personal data is shared, while creating a standardised infrastructure for data exchange.

Taken together, these initiatives illustrate the gradual development of a platform layer within India’s emerging data governance framework. Some initiatives focus on publishing government datasets; others enable consent-based data sharing between institutions; still others aim to build repositories for AI-relevant datasets. Newer efforts such as AIKosh seek to create national repositories of datasets and models for AI research, while language initiatives such as Bhashini and BharatGen reflect investments in multilingual AI infrastructure. These platforms have increased the visibility and availability of datasets that were previously scattered across ministries and institutional silos.

Despite these developments, the ecosystem remains fragmented, with multiple platforms operating in parallel without sufficient integration. While data availability has improved, the coexistence of these platforms does not automatically translate into usable, interoperable datasets. In the absence of shared standards and coordination across systems, datasets remain uneven in quality, documentation, and accessibility.

The Institutional Layer

MeitY’s draft National Data Governance Framework Policy proposes the India Data Management Office (IDMO) as India’s nodal non-personal data regulator to be established under the Digital India Corporation. The IDMO would be responsible for framing, managing, and periodically reviewing the National Data Governance Framework Policy, while developing the rules, standards, and guidelines required to implement it. Its proposed functions include formulating data, dataset, and metadata rules in consultation with ministries, state governments, and industry; developing anonymisation standards to open up data held by both government and non-government entities; and creating and managing the India Datasets Platform to enable researchers and startups to access anonymised datasets for AI development. However, the IDMO remains at the planning stage. The draft NDGFP has been under deliberation since 2022, and no legal mandate has yet followed. Until the policy is finalised and the IDMO receives statutory backing, its proposed functions remain aspirational.

Recommendations

The following recommendations should be read as mutually reinforcing institutional reforms rather than isolated interventions. Each addresses a different aspect of the data ecosystem, recognising that progress in one area depends on progress in the others. Together, they work to create conditions for data to be responsibly governed, effectively shared and reused, and ultimately translated into public value.

Figure 3: Recommendations for building responsible data infrastructure.

Invest in Responsible Data Infrastructure as a Strategic Pillar

Data infrastructure should be treated as a core public priority of India’s AI ecosystem, alongside compute and talent. Without sustained investment in data systems, AI development will remain constrained by poor data quality and fragmentation. MeitY should therefore position responsible data infrastructure as a central pillar of the IndiaAI Mission, with dedicated funding for data curation, standardisation, documentation, and maintenance. This includes building interoperable platforms, secure data exchange systems, and governance frameworks that ensure data is reliable and usable across sectors. It should also extend beyond technical platforms to the institutional capacity that keeps data usable, including the data skills of the departments and data holders responsible for producing and maintaining it. International experience shows that becoming a data-driven public sector requires continuous investment in data quality and governance, rather than one-time digitisation efforts. Prioritising data infrastructure in this way is essential for developing AI systems that are reliable, accountable, and effective in practice.

Create AI-Ready Data Standards

National AI-ready data standards are now essential for reliable public-sector AI deployment. MeitY should therefore establish and enforce national standards for AI-ready datasets used in public services. The UK government’s Guidelines for AI-ready datasets emphasise the importance of structured metadata, clear documentation, and interoperability across systems. Similarly, OECD (Organisation for Economic Co-operation and Development) guidance highlights the need for traceability, transparency, and bias mitigation in AI systems. Principles such as FAIR (Findable, Accessible, Interoperable, Reusable) also stress that data should be usable and reusable across contexts. Open data standards extend the same principle by making datasets publicly available in standardised, machine-readable, and openly licensed formats. 

India’s AI-ready standards should go beyond basic data quality checks. They should define requirements across three areas. First, metadata and documentation, including standardised dataset descriptions, coverage, and known limitations. Second, data quality, including accuracy, completeness, consistency, and representativeness. Third, provenance, including clear records of how data was collected, processed, and updated over time. Compliance should progressively become mandatory. Periodic audits and public documentation should also be required for datasets used in high-impact applications. These standards would support the proactive publication of high-value datasets in areas such as public health, agriculture, land, and public finance, where better access can enable research, innovation, and public-interest AI. Since standards are only as effective as departments’ ability to meet them, national data platforms should also offer in-house technical support to help data holders structure, document, and publish their datasets to these standards.

Incentivise Inter-Departmental and Public-Private Data Sharing

MeitY should establish structured data-sharing partnerships across government and with the private sector, moving beyond voluntary or ad hoc approaches. While existing initiatives such as the NDAP, AIKosh, and the Draft India Data Accessibility and Use Policy have laid important foundations for data sharing, their implementation has remained uneven across departments. This reflects not only the absence of incentives, but also capacity constraints, fragmented ownership of datasets, and limited institutional support for preparing data for reuse. Data-sharing arrangements should therefore combine incentives with common standards, clear governance arrangements, and technical support to reduce the administrative burden associated with sharing high-quality datasets. This can include linking departmental performance metrics to the sharing of high-quality datasets, mandating the timely release of non-sensitive, high-value public data, and providing contributing departments with priority access to shared AI infrastructure and national data platforms. Existing policy frameworks should be strengthened through enforceable standards, regular updates, and interoperability requirements to ensure that shared data is usable for AI applications. They should also provide organisations with confidence that shared data will be managed transparently, used consistently with agreed purposes, and subject to appropriate accountability mechanisms.

For public–private data exchange, MeitY should develop incentive-compatible partnership models that combine clear benefits with risk safeguards. These partnerships should be supported through standardised data-sharing agreements, licensing frameworks, and governance protocols that define rights, responsibilities, and permitted uses. Approaches such as the Data Empowerment and Protection Architecture (DEPA) demonstrate how trusted, auditable data-sharing systems can enable participation while maintaining accountability. Building on this, private entities can be drawn in through regulatory clarity, co-investment opportunities, and preferential access to public datasets or compute infrastructure, while initiatives such as AIKosh can serve as anchor platforms for such exchange. Embedding these mechanisms will help make data sharing a routine institutional practice rather than a discretionary activity.

Build Domain-Specific Data Collaboratives 

India should establish domain data collaboratives in health, justice, agriculture, climate, and urban governance. These collaboratives should be structured around shared, multi-stakeholder governance that brings government agencies, civil society organisations, research institutions, and private sector data holders together to produce and maintain AI-ready datasets in their sectors. The domain collaborative model addresses the incentive problem that voluntary frameworks cannot solve. When multiple actors share both the costs and governance of a common data resource, and when outputs are available to all contributors, the incentives to participate shift. Contributors gain access to integrated datasets that no single institution could assemble alone, while retaining a role in shaping how the data is used.

International experience shows that such models can deliver real public value. Data collaboratives have been used in areas such as climate analysis, where public and private actors combine datasets to improve decision-making and coordination. For example, during the Ebola outbreak in Nigeria, public authorities worked with telecom and health organisations to share data and track disease spread in real time. These collaboratives also enable new forms of value creation, including improved public service delivery, better forecasting, and more effective policy evaluation. In India, initiatives such as the India Urban Data Exchange and state efforts like Telangana Data Exchange indicate the potential for shared data infrastructure, but these need to be expanded into formal collaboratives with clear governance frameworks, incentives, and interoperability standards. Establishing such domain-level institutions will be critical to enabling cross-sectoral AI applications and ensuring that data sharing produces collective value. 

Institutionalise Data Stewardship Roles within Government, Academia, and Civil Society

Data stewardship should be treated as a core institutional function across government, academia, and civil society. At its core, stewardship is about making data usable without undermining the agency of the people and communities from which it is generated. Stewards can act as trusted intermediaries between data holders and users, helping curate and standardise datasets while ensuring that consent, accountability, and community participation are built into how data is used. This also helps address deeper concerns around bias, representation, and trust. 

A concrete institutional form for this is the data trust model. The Vidhi Centre for Legal Policy argues that stewardship through data trusts is better suited to community-centric data governance than either open data sharing or direct commercial exchange. It proposes a framework in which a board of trustees, supported by sub-committees representing data providers, users, and community members, manages how non-personal data is accessed and used. The model builds in enforceable fiduciary duties, transparency requirements, including public disclosure of policies and annual reports and accessible grievance mechanisms. Such arrangements can strengthen accountability and provide institutional mechanisms for balancing competing interests in data governance.

At the same time, important questions about who legitimately represents affected communities, how consent and community preferences are maintained as data is reused over time, how conflicts between public interest and commercial incentives are resolved, and how trustees themselves are held accountable. Stewardship mechanisms alone may not overcome wider structural constraints such as weak public data infrastructure or unequal bargaining power between large technology firms and smaller public institutions. However, stewardship mechanisms will only succeed if they are embedded within a broader governance framework that combines clear legal standards, institutional capacity, transparency, and public oversight. 

Concluding Remarks

India’s AI ambitions will depend not only on advances in compute infrastructure, foundation models, or applications, but also on the quality and governance of the data that trains them. This essay has argued that many of the barriers to responsible AI development arise before model training begins. Fragmented datasets, inconsistent data quality, limited interoperability, weak governance, and inadequate mechanisms for representation and consent undermine the reliability of AI systems. They also limit who is able to participate in India’s AI ecosystem and who ultimately benefits from it.

A coordinated data ecosystem should therefore not be understood simply as a way to make AI development faster or more efficient. Its purpose is to ensure that data is governed in ways that support trustworthy, accountable, and socially beneficial AI. This requires improving the quality and accessibility of datasets. It also requires enabling responsible data sharing across institutions, strengthening transparency and independent scrutiny, and ensuring that AI systems better reflect India’s linguistic, social, and regional diversity. Public data infrastructure is valuable not only because it enables innovation, but because it creates the conditions for better public services, evidence-based policymaking, and wider participation in AI development.

The recommendations proposed in this essay are intended to strengthen these institutional foundations rather than prescribe a single governance model. Building responsible data infrastructure is a continuous process that requires coordination across government, academia, industry, and civil society. As India continues to expand its AI capabilities, equal attention should be given to the institutions that govern data. A stronger data ecosystem is therefore not simply an input to AI innovation; it is a prerequisite for ensuring that India’s AI ecosystem is trustworthy, inclusive, and capable of delivering lasting public value.