AI Data

Overview

AI data in India sits at a pivotal inflection point: the country possesses extraordinary raw assets—one of the world's most linguistically diverse populations, a mature digital public infrastructure stack, and roughly 40% of the global AI engineering workforce—yet the governance frameworks, stewardship practices, and equitable distribution mechanisms needed to convert those assets into sovereign AI capability remain dangerously incomplete 1. The bottleneck is not compute or model sophistication; it is the absence of trustworthy data pipelines, clear monetization and consent frameworks, and representative datasets that reflect India's social and linguistic reality. Meanwhile, the window to act is narrowing: path dependency in global AI development, accelerating model capabilities, and consolidating market power among frontier labs mean that the terms on which Indian data enters the AI economy are being set now, often unfavorably 23. Getting this right matters not only for India's own citizens—particularly the hundreds of millions who are non-English-speaking, rural, or digitally marginal—but for the Global South broadly, which increasingly looks to India's digital public infrastructure as a replicable model.

---

Key Insights

• Data governance is a more urgent bottleneck than compute. Across sectors—health, agriculture, language, energy—the binding constraint is not GPU availability but the absence of clear DPDP rules for AI training, data monetization frameworks, and interoperability standards. Enterprises have data; they cannot feed it to AI because governance, lineage, metadata, and access controls are broken 14.

• Federated architectures are the practical path to data sharing without data surrender. Federated learning, clean-room environments, and data-visitation models (send algorithms to data, not data to cloud) are no longer experimental—they have demonstrated proof-of-concept in genomics 5, healthcare benchmarking via BODH 6, and ocean science 7. The pattern is consistent: local data control plus shared intelligence is technically feasible today and should be the default design choice for sensitive domains.

• Metadata, glossaries, and provenance are as important as the data itself. Without machine-readable definitions, data remains siloed even when digitized. Three technical prerequisites for AI-ready data recur consistently: enriched metadata, interoperable standards, and provenance documentation 89. India's National Statistical Offices and sectoral data custodians need investment in this layer before more model-building.

• Community agency—not just community inclusion—determines whether data systems are equitable. The structural distinction between communities as data providers versus data owners and controllers is fundamental 23. Frameworks that consult communities without giving them power over problem definition, curation decisions, and benefit distribution replicate extractive dynamics regardless of their stated intent 1011.

• India's manuscript and oral heritage represents a finite, authentic, and globally unique data asset. The Gyan Bharatam initiative demonstrates that culturally-grounded training corpora—built from OCR of ancient manuscripts across scripts and languages—can create genuine competitive advantage in building AI systems that reflect Indian knowledge systems rather than Western proxies 12. This asset is irreplaceable and time-sensitive.

• Government datasets are under-utilized national infrastructure. Broadcasting archives, agricultural records, road safety databases like IRAD/EDAR 13, health records within ABDM 6, and ocean observation data from the Ministry of Earth Sciences 7 are all identified as ready-to-mobilize assets blocked by legal ambiguity, organizational incentives to hoard, and the absence of API-enabled access layers—not by technical limitations.

• Data representation is a political and rights question, not merely a technical one. The decision to include or exclude populations—women, lower-caste communities, non-Hindi speakers, tribal groups—from training datasets is not a neutral modeling choice. Incomplete farmer registries produce algorithmic exclusion 14; the ALIGN Benchmark's engagement of 20,000 women across six Indian languages demonstrates that participatory data collection produces measurably better, more culturally valid systems 11.

• Open standards and interoperability are the antidote to extractive lock-in. Repeatedly, the analogy to UPI recurs: just as India leapfrogged legacy banking infrastructure through open payment rails, solving data standardization and interoperability now can position India to lead in trustworthy AI 1516. MCP protocols, FAIR/CARE principles in genomics 5, and DPI "highways" models in agriculture 17 all reflect the same architectural logic.

• The economics of data contribution are broken and must be redesigned. Big tech companies use grassroots AI datasets—from community NLP projects, citizen science, and public-sector digitization—without consent or compensation 1018. Frameworks for fair data economics, including tiered marketplaces with contributor-controlled licensing and explicit benefit-sharing agreements, are technically feasible but require political will to mandate 319.

---

Recurring Themes

• Voice and vernacular as non-negotiable equity infrastructure. Nearly every sector brief—health 20, agriculture 17, education 21, governance 15—independently converged on the same finding: text-based, English-first AI excludes the majority of India's population. Bhashini's 22-language capability 22, IVR-based agricultural advisory systems 17, and voice-based health interfaces 20 represent not feature additions but foundational equity requirements. The engineering community treats this as solved; the deployment gap remains vast.

• The sequencing problem: governance must precede or accompany data deployment, not follow it. From healthcare (ABDM built before AI rollout 6) to education (policy frameworks before technology rollout 23) to open data (regulation required for sustainability 24), speakers across domains independently argued that voluntary initiatives fail at scale and that starting with open, non-sensitive data to build trust—then layering in sensitive data with stricter controls—is more effective than attempting full governance from the outset 25.

• Data as colonialism: extraction without redistribution is a recurring structural critique. Whether framed through labor rights in AI supply chains 18, indigenous language erasure through LLMs 2, women's knowledge extraction without control redistribution 11, or community datasets scraped by frontier labs 10, a consistent structural argument emerges: the current data economy concentrates value at the frontier while externalizing costs onto the Global South. This is not a peripheral concern—it was raised across tracks as a defining governance failure of the current AI moment.

• Hyperlocal, disaggregated data is systematically missing. Climate resilience 2627, agricultural advisory 28, disaster management 29, and road safety 13 all identified the same gap: global and national models fail because they lack neighborhood- or farm-plot-level ground truth. The solution is not more centralized data collection but sustained investment in community-level data generation, with appropriate governance and incentive structures.

• Open-source tooling for evaluation and safety is as urgent as open-source models. Multiple speakers—covering healthcare AI assessment 30, education benchmarking 21, general AI assurance 31, and open-source safety tools 32—converged on the point that the evaluation and benchmarking layer is under-resourced relative to model development. Without accessible, contextually adapted evaluation frameworks, responsible procurement and deployment are impossible, particularly for under-resourced institutions in the Global South 33.

---

Open Challenges & Tensions

• DPDP rules for AI remain unresolved, creating a governance vacuum with real costs. The Digital Personal Data Protection Act has not clarified pathways for federated learning, synthetic data validation, cross-institutional sharing, or AI training on public datasets 13435. This ambiguity simultaneously chills legitimate innovation (hospital data collaborations, genomic research) and fails to constrain extractive practices by actors willing to operate in legal grey zones. The tension between enabling innovation and protecting rights has not been operationalized—it remains a set of aspirational principles.

• Sovereignty versus openness is a genuine and unresolved tension. The case for data sovereignty—territorial control over sensitive datasets, local computation, protection from asymmetric extraction—pulls against the equally compelling case for open, interoperable global commons that accelerate progress for all. Bhashini's migration from hyperscale cloud to locally controlled infrastructure while maintaining Nvidia and Azure partnerships 36 illustrates one pragmatic middle path, but the terms of that balance have not been settled at the policy level, and differ materially depending on whether the actor is a G20 government or a low-income community.

• Certification and standards risk measuring the measurable, not the meaningful. AI assurance frameworks, data quality certifications, and responsible AI standards were endorsed across the summit—but a consistent counter-argument also emerged: quantifiable metrics (transcription accuracy, dataset size, benchmark scores) systematically under-represent dimensions of worker dignity 18, cultural validity 10, and community agency 192 that resist standardization. The field has not resolved how to design accountability systems that are rigorous without being reductive.

• Speed of deployment versus readiness of governance infrastructure. The commercial and political incentive to deploy AI at scale is running significantly ahead of the governance, data quality, and institutional capacity needed to do so safely. In healthcare, only 45% of government leaders plan to evaluate pilots 37; in education, contextual validation is treated as optional rather than prerequisite 21; in agriculture, pilots accumulate without systematic documentation of failure 28. The summit produced broad consensus that evaluation must precede scale—but no mechanism for enforcing that sequencing against deployment pressure.

• Who speaks for communities in data governance remains structurally unclear. Participatory frameworks, multi-stakeholder consultations, and community co-design were universally endorsed—but the operational question of which stakeholders engage at which lifecycle stages, with what binding authority 2, was rarely answered. The risk is that participation becomes performance: communities are consulted on decisions already made, extractive dynamics are relabeled as co-creation, and the structural power imbalance between frontier AI developers and data-source communities is unchanged.

---

Notable Examples

• BODH (Benchmarking Open Data Platform for Health AI), India's federated health AI benchmarking initiative 6: Sends algorithms to data rather than data to cloud, solving both fragmentation and privacy concerns for Indian healthcare AI validation. Cited as a model for how governance can enable rather than block innovation—potentially replicable in agricultural and education AI.

• Bhashini / Pashini platform migration 2236: India's 22-language government AI platform, originally hosted on hyperscale cloud, was successfully migrated to locally controlled infrastructure while maintaining partnerships with Nvidia and Azure. Described as proof that middle-income nations can build sovereign AI stacks without autarky—a direct counter to the false binary between openness and control.

• Gyan Bharatam manuscript digitization initiative 12: Combines OCR, LLMs, speech technology, and semantic search to extract, authenticate, and democratize knowledge from ancient Indian manuscript corpora across scripts and languages. Explicitly framed as creating training data that encodes Indian knowledge systems—a finite, authentic competitive asset unavailable to Western frontier labs.

• ALIGN Benchmark—20,000 women, six Indian languages 11: A participatory data collection and evaluation initiative engaging 20,000 women to define and identify gender bias across six Indian languages. Demonstrates that community-led benchmarking produces more culturally valid AI systems than English-centric approaches, and that intersectional representation in training data is methodologically, not just ethically, necessary.

• Mahavistar / Bharat Vistar DPI agricultural platform 14: Open, interoperable digital backbone for agriculture—farmer IDs, data exchanges, shared protocols—reaching 2.5 million users with planned national expansion as Bharat Vistar. Cited alongside FAO's federated digital public goods model 38 as evidence that DPI-first architecture, combined with deliberate inclusion of women and marginal farmers in data registries, enables both scale and equity in ways that proprietary platforms cannot.