Ankit Bose

Abstract

India stands at an inflection point in the global AI transition, positioned by the scale of its population, digital public infrastructure, and developmental ambitions. Yet the dominant AI paradigm, shaped by capital-intensive firms and concentrated technology supply chains, may reproduce structural inequities unless mediated by purposeful public policy. This essay advances a framework of Inclusive AI Sovereignty, defined not as technological autarky or unilateral withdrawal, but as the democratic capacity to set and enforce the terms on which AI affects Indian residents. The framework prioritises four linked tasks: a binding horizontal accountability baseline; public-interest compute, data, and multilingual infrastructure; institutions for worker transition; and strategic capability at critical bottlenecks combined with international interoperability. It also treats sovereignty as contestable, requiring enforceable rights, transparency, and participation.

I. The Asymmetric Promise of AI in India

Artificial intelligence is not a neutral technology. AI systems carry the imprint of the data, institutions, incentives, and social contexts in which they are designed and deployed. As India engages with the AI transition, the central policy question is not merely how rapidly adoption should occur, but for whom, on what terms, and under whose authority. This essay uses sovereignty in a limited institutional sense: the capacity of Indian democratic institutions to set, enforce, and revise the terms on which AI is developed and used. It does not imply self-sufficiency at every technical layer or disengagement from international cooperation. The relevant choice is therefore not simply between the domestic and the global. Rather, it is which capabilities must be held or credibly controlled in India, which dependencies can be diversified through trusted partnerships, and which rights must remain enforceable regardless of where a model or cloud service is supplied.

India’s encounter with AI is structurally asymmetric in at least four dimensions. First, the foundational models that power most AI applications today have been built overwhelmingly on data generated in the Global North, in English, and by a narrow segment of the global population. These models embed assumptions, norms, and linguistic structures that are often alien to the lived realities of India’s 1.4 billion people, who speak over 22 scheduled languages, belong to thousands of castes and communities, and inhabit vastly different socioeconomic geographies.

Second, access to advanced compute has expanded through IndiaAI’s shared facility, which was reported to exceed 45,000 Graphics Processing Units (GPUs) by June 2026. This materially lowers barriers to entry for startups, researchers, and public institutions. However, frontier accelerators, semiconductor manufacturing equipment, hyperscale cloud infrastructure, and many general-purpose models remain concentrated among a small number of foreign suppliers. Shared access does not by itself guarantee supply-chain resilience, workload portability, model auditability, bargaining power, or continuity of service. The dependency problem is therefore narrower, but more concrete, than an argument that all compute must be domestic.

Third, the labour-market consequences of AI are likely to be significant but should not be presented as predetermined. A NITI Aayog roadmap developed with Boston Consulting Group (BCG) and National Association of Software and Service Companies (NASSCOM) presents both a downside scenario of job displacement and an upside pathway in which India creates millions of AI-related opportunities. A survey by the Indian Council for Research on International Economic Relations (ICRIER) of 651 Indian IT firms found no evidence of large-scale net job losses, but did identify a broad moderation in hiring concentrated at the entry level. The ILO similarly cautions that occupational exposure signals potential transformation rather than inevitable replacement. For India, the immediate policy risk is the weakening of entry-level pathways in IT-enabled services, business-process operations, and routine software work that have historically enabled upward mobility. This justifies anticipatory transition policy without assuming an unavoidable employment collapse.

Fourth, and perhaps most consequentially for democratic governance, AI systems are being deployed in public administration, in welfare delivery, policing, border management, and judicial support with inadequate attention to their discriminatory potential. India’s social fabric is deeply marked by caste, religion, gender, and class. AI systems trained on historical data risk institutionalising and amplifying these biases at unprecedented scale and speed, converting what were once the judgments of fallible but accountable human officials into the opaque outputs of algorithmic systems.

Taken together, these asymmetries create a policy chain. Concentrated external capabilities can reduce the ability of Indian institutions to inspect or remedy failures; weak legal duties can transfer those failures to citizens; and unequal labour, linguistic, and social conditions determine who bears the cost. The recommendations in this essay correspond directly to that chain: enforceable accountability for high-stakes uses; public-interest infrastructure and meaningful linguistic inclusion; worker-transition institutions; and resilient domestic capability combined with international cooperation.

India’s Distinctive Context

Any serious AI policy for India must begin not with the templates of Silicon Valley or Brussels, but with India’s own realities. India is simultaneously the world’s largest democracy, one of its fastest-growing major economies, a country with extraordinary linguistic and cultural diversity, and one where hundreds of millions still lack reliable access to basic healthcare, quality education, and economic opportunity.

India’s digital infrastructure is both a remarkable achievement and a source of vulnerability. The India Stack, encompassing Aadhaar, UPI, DigiLocker, and the Account Aggregator framework, has created a powerful foundation for digital public infrastructure (DPI) that is the envy of many countries. India’s Jan Dhan-Aadhaar-Mobile (JAM) trinity has enabled direct benefit transfers that have transformed welfare delivery. These achievements are real and significant. But they also mean that the state now holds unprecedented data about the lives of ordinary Indians. This data, if coupled with powerful AI systems without adequate safeguards, creates conditions for surveillance, exclusion, and control at a scale previously unimaginable.

India’s AI policy must therefore grapple with a fundamental tension: how to harness the genuine potential of AI to accelerate development while ensuring that the power concentrated in AI systems remains accountable to democratic institutions and that the benefits are distributed equitably across the full diversity of Indian society.

II. Current Approaches and Gaps: From Ambition to an Emerging Architecture

India’s engagement with AI policy has intensified substantially since NITI Aayog released the National Strategy for Artificial Intelligence in 2018. By 2026, the IndiaAI Mission, the India AI Governance Guidelines, sector-regulator initiatives, and the outcomes of the India AI Impact Summit had begun to form an identifiable governance architecture. The issue is therefore no longer one of absence of ambition or activity. It is whether this emerging architecture acquires binding legal force, independent oversight, coherent institutional interfaces, and measurable public accountability.

National Initiatives: Achievements and Limitations

The IndiaAI Mission, approved in 2024 with an outlay of approximately INR 10,372 crore over five years, remains India’s most significant public AI programme. By June 2026, the Government reported a shared compute facility of more than 45,000 GPUs; AIKosh holdings of 12,519 datasets, 307 models, and 20 toolkits; and support for a portfolio of Indian foundational and small language models. These represent material advances in compute access, datasets, model development, applications, startup financing, skills, and safe-and-trusted AI capacity.

The central limitation is no longer captured adequately by comparing headline budgets with the United States, China, or the European Union. The sharper question is whether public investment changes access and control at strategic bottlenecks. IndiaAI should publish utilisation rates, waiting times, user and sector mix, supplier concentration, geographic allocation, energy use, and public-benefit outcomes; reserve capacity for universities and public-interest projects; and require portability and supplier-diversification plans. Without these measures, aggregate GPU counts can coexist with unequal access and continuing dependency.

The governance critique also requires greater precision. The India AI Governance Guidelines state that public feedback was considered and recommend an AI Governance Group, a Technology and Policy Expert Committee, and an AI Safety Institute. These are important steps. However, the publicly available arrangements do not yet provide statutory guarantees of balanced representation of labour, civil society, state governments, affected communities, and independent researchers; publication of minutes and conflicts of interest; or an independent channel for complaints about public-sector AI. The concern is therefore not that participation is wholly absent, but that it is not yet sufficiently institutionalised or enforceable. Critical scholarship on the 2026 Summit raises a related concern about the risk of public-interest language coexisting with concentrated state and corporate power.

India also now has non-binding national AI governance guidance and a growing body of sectoral instruments. Yet no comprehensive statute creates general rights to notice, meaningful explanation, human review, contestation, incident reporting, and independent audit across high-risk automated decisions. The Digital Personal Data Protection Act, 2023, governs personal-data processing but does not cover the full lifecycle of AI systems, non-personal model risks, or every form of algorithmic accountability. The remaining gap is therefore between principles and enforceable rights, not between complete inaction and a single omnibus law.

Sectoral Deployments: Promise and Peril

AI is already being deployed across multiple public-interest sectors, although the maturity and evidence base differ. Kisan e-Mitra operates in 11 languages and had responded to more than 95 lakh queries by early 2026. IndiaAI’s 2026 health compendium documents use cases including AI-assisted diabetic-retinopathy screening and other diagnostic or workflow applications. In the justice system, LegRAA, SUPACE, ASR-SHRUTI, and PANINI are being developed or piloted as research, transcription, and translation aids, while preserving judicial control. These examples demonstrate real deployment; they do not establish uniform effectiveness. Independent evaluation of accuracy, distributional impact, security, and user redress remains necessary.

The common thread is deployment moving faster than enforceable governance in some high-stakes settings. The relevant risks are concrete: wrongful exclusion from benefits; biased triage or prioritisation; hallucinated medical, legal, or administrative advice; failures in dialects and code-switched speech; privacy and cybersecurity breaches; undisclosed model changes; and the inability of an affected person to understand or contest an output. Structural concentration can exacerbate these risks. An Indian deploying agency may have limited access to model weights, training data, evaluation methods, safety incidents, or the supplier’s upgrade cycle, while guardrails designed for dominant languages and markets may miss caste, religious, regional, and domain-specific harms. In high-stakes services, technical dependency can therefore become a due-process and accountability failure.

The Missing Dimensions

Labour policy remains a major gap, but the evidence should be stated carefully. The NITI Aayog roadmap identifies both displacement risks and a potential pathway to large-scale job creation, while ICRIER finds no current aggregate employment collapse but a measurable moderation in hiring concentrated in entry-level roles. This supports anticipatory transition policy rather than a deterministic mass-unemployment claim. Existing skilling initiatives are not yet sufficiently integrated with granular labour-market monitoring, employer co-financing, redeployment commitments, income support, and portable social protection for workers whose tasks change faster than formal qualifications.

AI already supports legal research, translation, transcription, and case-management experimentation in Indian courts through systems such as LegRAA, SUPACE, ASR-SHRUTI, and PANINI. In 2026, a preliminary Draft Regulations for Use of Artificial Intelligence in Courts proposed principles including human primacy, transparency, accountability, data protection, and judicial independence. The remaining gap is the transition from pilots and draft norms to enforceable, court-wide rules governing disclosure, verification, procurement, audit access, data handling, record retention, and the ability of parties to challenge AI-influenced material.

India’s international posture also requires updating. The India AI Impact Summit 2026 produced a declaration endorsed by 92 countries and international organisations, commitments signed by 13 model providers, an Equitable AI Transition Playbook, and the proposed Trusted AI Commons with 23 partner countries. One frontier commitment calls for multilingual and contextual evaluations based on datasets, benchmarks, and expertise for under-represented languages and cultures. These outcomes demonstrate agenda-setting capacity rather than a merely reactive posture. The unresolved issue is implementation: many commitments are voluntary and non-binding, and India does not yet have a unified domestic mechanism to publish benchmark results, compliance progress, investment additionality, and the distribution of benefits. Strategic leadership therefore means converting convening outcomes into measurable domestic rules and durable shared infrastructure.

III. Recommendations: Four Priorities for Inclusive AI Sovereignty

India needs an AI compact: a shared and enforceable understanding among the state, industry, workers, civil society, researchers, and citizens about the purposes AI should serve, the rights it must respect, and the accountability it must bear. To avoid an undifferentiated catalogue of proposals, the recommendations are organised around four priorities. First, establish rights and accountability for high-stakes uses. Second, build public-interest infrastructure and meaningful linguistic inclusion. Third, create institutions for worker transition and the equitable distribution of risk. Fourth, secure strategic resilience through managed interdependence and international leadership. This ordering separates immediate legal protections from medium-term infrastructure and long-term industrial policy.

Pillar I: Democratic Governance and Accountability

The first priority is a hybrid governance architecture rather than an exclusive choice between a single horizontal AI regulator and fragmented sectoral rules. Parliament should enact a statutory, technology-neutral baseline defining high-risk uses, prohibited practices, lifecycle duties, and individual rights. Sector regulators, courts, and relevant state authorities should translate that baseline into binding domain rules. A national safety and coordination institution should maintain common technical standards, an incident registry, and cross-sector oversight. This preserves consistent rights while using the expertise of institutions that understand healthcare, finance, employment, education, welfare, policing, and justice.

• A risk-tiered statutory baseline that subjects high-risk applications in criminal justice, welfare administration, healthcare, education, employment, and essential financial services to pre-deployment impact assessments, conformity testing, documented human oversight, and ongoing monitoring.
• Mandatory notice and transparency for AI used in public administration, including disclosure whenever AI materially influences a decision affecting a person’s rights or access to services, reasons intelligible to the affected person, and a practical route to human review and appeal.
• Prohibitions or tightly drawn statutory restrictions for uses presenting unacceptable risks to fundamental rights, including social scoring by public authorities, manipulative systems that exploit vulnerability, and real-time biometric surveillance in public spaces except under narrowly defined law, necessity, proportionality, and independent authorisation.
• Specific protections for AI in the justice system, including judicial authorisation where appropriate, disclosure to parties, verification duties, preservation of the record, and exclusion or reduced evidentiary weight where an AI output cannot be meaningfully examined and challenged.
• A statutory AI Safety and Accountability Office, institutionally separated from promotional functions, to issue common standards, accredit auditors, maintain a national incident register, investigate cross-sector failures, and coordinate enforcement. Sector regulators would retain primary responsibility within their domains rather than being displaced by a single all-purpose regulator.

A purely horizontal regime can provide uniform rights but may lack the contextual expertise needed to evaluate clinical safety, lending discrimination, judicial process, or labour practices. A purely sectoral regime benefits from domain knowledge but may create gaps, inconsistent definitions, duplicative compliance, and regulatory arbitrage. Comparative governance frameworks increasingly combine technology-neutral principles and risk classification with sector-specific implementation. India’s own AI Governance Guidelines also adopt a whole-of-government approach while recognising sectoral responsibility. A hybrid model is therefore better suited to India’s federal and regulatory structure than either extreme.

Democratic oversight must accompany regulatory administration. Parliament should establish a standing mechanism to scrutinise major government AI deployments, commission independent assessments, and review annual incident and compliance reports. State legislatures and state-level oversight bodies need equivalent capacity because policing, welfare, education, public health, and local administration frequently sit within state competence. Oversight should focus on outcomes, rights, and public expenditure rather than attempting to micromanage technical design.

Pillar II: Equitable Distribution of Benefits and Risks

India’s AI strategy must be animated by an explicit commitment to ensuring that the benefits of AI reach marginalised communities and that the risks of displacement, discrimination, surveillance, and administrative error do not fall disproportionately on those least able to bear them.

The principal public lever is a National AI for Public Goods Programme that builds on, rather than duplicates, the IndiaAI Mission. It should prioritise a limited portfolio of high-value use cases in primary healthcare, foundational learning, agricultural extension, disability access, and social-protection administration. Funding should be tied to clearly defined public outcomes, independent evaluation, open and interoperable components where feasible, procurement transparency, security testing, and continued human alternatives for people who cannot or should not use an AI channel.

India’s Digital Public Infrastructure experience is relevant because common rails, open interfaces, shared standards, and clear institutional roles can lower entry costs and enable multiple providers to innovate. But an API or repository does not become a public good merely because it is widely accessible. AIKosh already provides datasets, models, toolkits, use cases, and sandbox resources, while the 2026 Summit announced a voluntary Trusted AI Commons. A public-interest AI layer should build on these platforms and add enforceable safeguards: model and dataset cards; provenance, consent, and licensing records; representative quality tests; independent security and bias audits; tiered access for sensitive assets; transparent public-interest compute quotas; procurement disclosure; and mechanisms for correction, withdrawal, complaint, and benefit sharing. The DPI analogy is therefore institutional and governance-based, not simply a call for more open APIs.

Linguistic inclusion must be assessed as usability and social representation, not as a language-coverage checkbox. BHASHINI is an important national capability, but a claim that a system “supports” a language may conceal uneven performance across dialects, oral registers, code-switching, literacy levels, disability contexts, and domains such as health, agriculture, welfare, and law. Publicly available comparable evidence on these dimensions remains uneven. The 2026 frontier commitments appropriately recognise the need for multilingual and contextual datasets and benchmarks. A staged theory of change is preferable to an arbitrary deadline. By the end of 2027, high-volume central services should publish baseline performance, error categories, and complaint data and establish language and community advisory panels. During 2028–29, independent domain testing should extend across all 22 scheduled languages, prioritising high-stakes services. By 2030, a high-stakes public AI service should be deployed in a language only when it meets published accuracy, safety, and usability thresholds; offers a reliable human alternative; and provides local redress. Evaluation should combine technical metrics with participatory testing by native speakers, dialect communities, domain professionals, persons with disabilities, and users with low literacy.        

A further priority is a just and evidence-led transition for work. India should establish a National Commission on AI and Employment to maintain task-level and sector-level indicators, publish regular vulnerability assessments, coordinate employer and worker evidence, and recommend time-bound responses. The Commission should distinguish productivity augmentation from actual displacement and should evaluate whether new jobs provide comparable wages, security, and geographic access.

• Reskilling linked to employment outcomes. The National Skill Development Corporation, Skill India Digital, and Pradhan Mantri Kaushal Vikas Yojana (PMKVY) should move beyond short, certification-led courses toward modular learning tied to verified employer demand, apprenticeships, redeployment, and wage progression. Employers receiving public incentives for AI adoption should co-finance transition pathways for affected workers.
• Portable protection for platform, gig, and contract workers. Implementation of the Code on Social Security should be accelerated, with interoperable contribution records and benefits that follow workers across platforms and employers. Eligibility should be based on economic dependence and work conditions rather than narrow labels chosen by the contracting platform.
• Targeted transition mechanisms for concentrated disruption. Where displacement is rapid and geographically or occupationally concentrated, including selected BPO and entry-level technology roles, government and industry should use time-bound adjustment funds, career-navigation services, income support, and local economic diversification. The objective is not to preserve every existing task, but to prevent the cost of technological transition from being transferred entirely to workers and communities.

High-risk public-sector AI should be subject to disaggregated impact assessments across relevant dimensions such as caste, gender, disability, age, language, geography, and socioeconomic status, where collection and use of such information are lawful and ethically justified. Independent auditors should test both model performance and the institutional process surrounding it. Results, limitations, mitigation plans, and residual risks should be published in an accessible form. The Right to Information framework should cover public AI systems, subject only to narrow and reviewable exceptions for legitimate security and confidential commercial information.

Pillar III: Strategic Autonomy Through Managed Interdependence

Strategic autonomy is managed interdependence, not autarky. India cannot and need not reproduce every component of a globally distributed AI value chain. It does need assured access to critical resources, the ability to switch suppliers, control over sensitive public workloads, domestic competence at selected bottlenecks, and legal authority over systems that affect people in India. The following priorities make that concept operational.

• Assured compute access and resilience. India has already established shared access to more than 45,000 GPUs. The next step is not undifferentiated expansion. The Government should reserve a transparent public-interest quota; publish utilisation, waiting-time, allocation, and outcome metrics; diversify suppliers and regions; require workload portability and open interfaces; maintain a secure sovereign cluster for sensitive public workloads; and pre-negotiate contingency access and reciprocal research arrangements with trusted international partners. These arrangements should specify service continuity, audit access, cybersecurity obligations, export-control contingencies, data handling, and exit rights.
• Capability at strategic bottlenecks. The India Semiconductor Mission had approved 12 fabrication or packaging projects and 24 semiconductor-design projects by May 2026. The next phase should set measurable milestones for advanced packaging, memory and interconnects, chip-design intellectual property, compiler and kernel libraries, testing and verification, energy-efficient data centres, and specialised talent. Procurement and R&D support should be tied to tape-outs, yields, domestic design ownership, reusable software, and supplier diversification. Attempting leading-edge self-sufficiency across every chip category would be costly and unrealistic; India should select bottlenecks where capability creates durable leverage.
• Targeted data sovereignty, not blanket localisation. The Reserve Bank of India’s payment-data rule illustrates hard localisation for a defined regulated category. AI policy should reserve on-soil storage or processing for narrowly specified critical public-sector datasets and highly sensitive workloads, while allowing conditional cross-border flows elsewhere. Conditions should include enforceable regulator access, encryption and key control, audit logs, portability and deletion rights, breach liability, and safeguards in the destination jurisdiction. Public-interest datasets should include access and benefit-sharing rules. This differentiated approach protects legitimate sovereign interests without imposing blanket costs on research, trade, and routine business operations.
• AIKosh as a governed commons. IndiaAI already operates AIKosh as a hub for datasets, models, tools, use cases, and sandbox resources. The policy question is, therefore, what existing dataset efforts do not yet guarantee. Admission and continued hosting should require machine-readable provenance, licensing, consent and lawful-basis records; coverage and limitation statements; dataset and model cards; independent quality and security review; community challenge and correction; access tiers; withdrawal and redress procedures; and documentation of how public and private contributors share value and responsibility. Representative Indian data is not created by scale alone; it depends on governance, documentation, and meaningful participation.
• International implementation and normative leadership. India should use the institutions created around the 2026 Summit to convert voluntary commitments into annual public scorecards covering multilingual and contextual evaluation, safety reporting, labour transition, compute access, and delivery of announced investments. It should co-fund open benchmarks and incident-exchange mechanisms with Global South partners, promote interoperable standards that remain locally enforceable, and use trade and research partnerships to diversify dependencies rather than replace one dominant supplier with another.

A credible Indian “third way” cannot be a rhetorical midpoint between United States market power and state-centric control elsewhere. Emerging scholarship warns that public-interest language may coexist with closed systems, voluntary self-regulation, “openwashing”, surveillance, and continued dependence on a small group of technology firms. The same risk applies domestically: sovereignty can centralise power in the state or incumbent firms unless people can inspect, contest, and influence systems. The test of a third way is therefore distributive and institutional: who can audit and appeal, who captures economic value, who bears labour and environmental costs, and whether affected communities can refuse or redesign a deployment.

Pillar IV: Building the Institutional Infrastructure for the Long Term

All four priorities depend on institutional capacity. The immediate need is not another broad strategy document but a small set of durable institutions with clear mandates, technical competence, independent scrutiny, and public reporting.

• Coordinated and independent expertise. The proposed AI Governance Group, expert committee, safety institution, sector regulators, standards bodies, and data-protection authorities should have published mandates, interfaces, budgets, conflict-of-interest rules, and reporting duties. Public support for external expertise should be provided through arm’s-length competitive grants administered by independent research councils, university consortia, or foundations with mixed funding and disclosure requirements. It should not take the form of discretionary state financing for selected advocacy organisations. This allows civil society and academic institutions to build technical capacity without compromising their ability to criticise government or industry.
• Evaluation and incident infrastructure. India should build accredited testing laboratories, red-teaming capability, secure arrangements for regulatory access to models, multilingual and culturally grounded benchmarks, domain-specific evaluation suites, and a national incident database. Evaluation should be connected to public procurement, sectoral licensing, periodic reassessment, and recall or suspension powers when material risks cannot be mitigated.
• Participatory accountability. High-risk public deployments should require consultation with affected-community panels before procurement or scale-up. Agencies should publish impact assessments, intelligible data and model summaries, the feedback received, and reasons for accepting or rejecting material proposals. Complaints, human review, and appeals must be accessible in relevant languages and through non-digital channels. The consultation used in developing the India AI Governance Guidelines is a useful starting point; participation should become recurring, representative, and linked to actual decisions.

Conclusion: Toward an Indian AI Compact

The choices India makes about AI in the next five years will shape its economic and democratic trajectory for a generation. The stakes are high, and the risks of getting it wrong are real: an India that allows AI to deepen its inequalities, erode its democratic institutions, and entrench its dependencies on foreign technology companies will be diminished, not strengthened, by the AI revolution.

The framework of Inclusive AI Sovereignty proposed in this essay is not a programme of withdrawal. It combines domestic capability with enforceable rights, selective industrial policy, and deep international cooperation. Its purpose is to ensure that interdependence remains a choice that India can manage, rather than a dependency that external suppliers or domestic incumbents can exploit.

This approach takes India’s strengths seriously: the scale of its digital public infrastructure, the vitality of its technology sector, the institutional reach of its federal democracy, and the linguistic and cultural diversity of its society. It also recognises the limits of state capacity and the danger that the language of sovereignty can legitimise surveillance or protectionism. Progress should therefore be measured by public outcomes, contestable decisions, supplier and model portability, worker transitions, multilingual usability, and independent evidence rather than through expenditure or model counts alone.

What it requires is the political will to enact meaningful governance frameworks even when powerful interests resist them, the institutional capacity to implement and enforce those frameworks over time, and the democratic imagination to ensure that the millions of Indians most affected by AI have a real voice in the choices being made on their behalf. These include migrant workers whose employment is threatened, welfare beneficiaries whose eligibility will be determined by algorithms, farmers whose decisions will be mediated by AI advisories, and students whose education will be shaped by AI tutors.

India has repeatedly shown that it can build institutions around transformative technologies while adapting them to its developmental context. The AI compact India needs is no less ambitious: democratic in authority, inclusive in distribution, technically credible, open to international cooperation, and accountable in practice.