The Role of Government as a Provider of Data for Artificial Intelligence

Why Government Data Is Critical for AI Development

Governments are among the most significant collectors, collators, and producers of data in the world. They hold nationally and sub-nationally representative datasets that provide deep insights into social dynamics—from school attendance and social protection usage to crime patterns and healthcare system performance. This data is generated through the provision of government services including civil registration, healthcare, education, business registration, policing, and national statistics exercises like censuses.

The Global Partnership on Artificial Intelligence (GPAI) report argues that this government-held data represents an essential foundation for developing AI tools that address social challenges and developmental priorities. When properly shared and governed, it can drive improvements in government service efficiency, close gaps in access to education and healthcare, and support evidence-based policymaking.

However, providing government data to AI developers must be undertaken responsibly. The report examines foundational principles including respect for human rights, privacy, consent, inclusivity, and ethical use. For organizations seeking to make complex policy research more accessible, interactive document experiences offer a compelling way to engage stakeholders with dense analytical content.

Four Models for Government Data Sharing

The report identifies four primary mechanisms through which government-owned data can be provided to AI developers, each with distinct governance implications and risk profiles.

Open Data is the most well-known model, with several governments leading initiatives to make public datasets freely accessible. While open data facilitates broad access and supports transparency, it raises concerns about data access justice in low and middle-income countries (LMICs). The report recommends adopting FAIR principles (Findable, Accessible, Interoperable, Reusable) alongside data solidarity models.

Contracts represent bilateral agreements between governments and AI developers. These cover data collection, access, control, processing scope, security measures, and limits on third-party disclosure. Contract-based sharing offers precision but requires sophisticated legal capacity and negotiation expertise that many governments lack.

Data Stewardships involve independent intermediaries that manage data access on behalf of governments and data subjects. This model has potential to grow, offering a balance between data utility and protection. The stewardship model can address power imbalances between large tech firms and government agencies.

Public-Private Partnerships enable collaborative arrangements where governments and private sector entities share governance responsibilities. While these can accelerate AI development, they require careful structuring to prevent conflicts of interest and ensure that public value is prioritized over commercial gain.

NHS and DeepMind: Lessons from a Controversial Partnership

Perhaps the most instructive case study in the report examines the UK’s National Health Service (NHS) data sharing arrangement with Google DeepMind. This case illustrates how even well-intentioned data sharing can go catastrophically wrong when transparency and governance are inadequate.

The NHS provided DeepMind with access to 1.6 million patient records from the Royal Free London NHS Foundation Trust to develop a clinical alert system called Streams. However, the arrangement was criticized for several reasons: the data sharing was not sufficiently transparent to patients, the scope of data accessed exceeded what was strictly necessary for the stated purpose, and the selection of Google DeepMind as the exclusive partner raised antitrust concerns.

The resulting public backlash included protests and investigations by the UK Information Commissioner’s Office, which found that the Royal Free had failed to comply with data protection law. This case demonstrates a fundamental lesson: government data sharing that bypasses public procurement processes and lacks transparency can cause lasting damage to citizen-government trust, ultimately undermining the very public services AI was meant to improve.

The report notes that governments must balance the urgency of AI innovation with the imperative of democratic accountability. Secret or opaque data sharing agreements, regardless of their technical merit, represent a governance failure that can set back AI adoption for years.

Taiwan’s Health Passbook: Citizen-Controlled Data Access

Taiwan’s Health Passbook offers a contrasting model to the NHS case—one where citizens maintain control over their own health data. The system allows individuals to access their personal health records through a secure digital platform and decide who can access that data and for what purposes.

The Health Passbook SDK architecture separates frontend operations (where users interact with their data) from backend file acquisition procedures. This technical design ensures that data remains under citizen control while still being available for authorized AI applications. The Taiwanese government provides the infrastructure and governance framework, but individuals retain agency over how their data is used.

This citizen-centric model represents an important alternative to top-down data sharing approaches. It demonstrates that government data provision for AI does not require centralized control over citizen data. Instead, it can work through empowering individuals to make informed decisions about data sharing, supported by robust technical infrastructure and clear legal frameworks.

The success of Taiwan’s approach depends on high levels of digital literacy and trust in government digital services—conditions that may not exist in all contexts but that offer an aspirational model for other countries.

Nigeria’s Rapid Response Register for AI-Driven Social Protection

The Nigerian Rapid Response Register (RRR) for cash transfers represents a different use case entirely—one focused on emergency social protection in a developing country context. The system uses government data to identify and reach vulnerable populations during crises, leveraging AI to improve targeting accuracy and reduce leakage in social protection programs.

This case study illustrates both the potential and the challenges of government data use for AI in LMICs. On the positive side, AI-driven targeting can significantly improve the efficiency of social protection spending, ensuring that limited resources reach those who need them most. The RRR demonstrates how government-held data—from civil registration, social welfare programs, and other administrative sources—can be combined and analyzed to support rapid response during emergencies.

However, the case also highlights concerns about data quality, consent, and the potential for exclusion errors in AI-driven targeting systems. In contexts where large segments of the population lack formal identification or digital presence, AI systems trained on government data may systematically exclude the most vulnerable—the very populations social protection is meant to serve. This creates a paradox that interactive analysis tools can help policymakers understand and communicate to stakeholders.

Colombia’s Aclimate: Agricultural Data for Climate Resilience

Colombia’s Aclimate platform demonstrates a successful model of government data provision for AI in the agricultural sector. Developed in partnership with the International Centre for Tropical Agriculture (CIAT) and Colombia’s National Institute of Hydrology, Meteorology and Environmental Studies (IDEAM), the platform combines government meteorological data with agricultural information to provide AI-powered agro-climatic forecasts to farmers.

The production and use of seasonal agro-climatic forecasts involves a sophisticated data pipeline: government agencies collect and process meteorological and hydrological data, which is then combined with agricultural datasets to generate actionable recommendations for farmers. This approach helps smallholder farmers make better planting, irrigation, and harvesting decisions in the face of increasing climate variability.

The Aclimate case is significant because it demonstrates how government data sharing for AI can create direct, tangible benefits for vulnerable populations. Unlike the NHS case, where the primary beneficiary was a private technology company, Aclimate channels AI capabilities directly toward improving the livelihoods of rural communities. The public-good orientation of the data sharing arrangement helps maintain public trust and ensures that the benefits of AI are distributed equitably.

Key Enablers for Responsible Government Data Sharing

The report identifies three essential enablers for effective government data sharing: a regulatory landscape ensuring equitable access, a policy framework facilitating data sharing, and infrastructure to support data accessibility.

On the legal side, five key areas demand attention. Antitrust regulation is critical because access to government data holds high economic value and can enable large AI developers to pursue anti-competitive practices that lock out smaller competitors. The UK case study illustrates this risk vividly, with Google DeepMind selected as the exclusive government partner.

Data protection law must evolve to address the specific challenges of government data sharing for AI, including purpose limitation, data minimization, and individual rights. Intellectual property frameworks need clarification regarding who owns insights derived from government data processed by AI systems. Freedom of information legislation can support transparency in data sharing arrangements. And sector-specific regulations—particularly in healthcare, finance, and social protection—must be adapted to the AI context.

Policy enablers include national AI strategies that explicitly address government data provision, data sharing guidelines for public servants, and institutional capacity building for data governance. Infrastructure requirements span digital connectivity, data storage capacity, standardized formats, and API-based access mechanisms that enable automated and scalable data sharing.

Mitigating Barriers and Privacy-Enhancing Technologies

Five common barriers emerge across the case studies: use of personal data to train AI systems, regulatory design challenges, value sharing disputes, technical capacity limitations, and adverse public attitudes. The report offers concrete mitigation strategies for each.

Privacy-enhancing technologies (PETs) represent the most promising technical solution. Recent developments have proven that data can be made available for AI training while protecting individual privacy through techniques including differential privacy (adding mathematical noise to prevent re-identification), federated learning (training AI models across distributed datasets without centralizing data), synthetic data generation (creating artificial datasets that preserve statistical properties), and homomorphic encryption (computing on encrypted data without decrypting it).

The report cautions that PETs alone are insufficient. As the OECD’s AI Policy Observatory confirms, they must be embedded within comprehensive governance frameworks that include legal protections, institutional oversight, and mechanisms for public accountability. Technology can support privacy but cannot substitute for democratic governance of data sharing decisions.

Value sharing remains a contentious issue. When government data contributes to commercially profitable AI products, questions arise about whether the public should benefit financially from that value creation. The report suggests that governments explore mechanisms including licensing fees, revenue sharing arrangements, and requirements for pro-bono public sector applications as conditions for data access. For teams working on these complex governance challenges, transforming policy documents into engaging formats helps build broader understanding and support.

Principles for Equitable AI Data Governance

The GPAI report concludes with recommended principles that synthesize lessons from all four case studies. These principles form a coherent framework for governments worldwide seeking to provide data for AI development while protecting citizens and maintaining public trust.

First, beneficial value must be demonstrable. AI systems utilizing government data should produce clear public benefits—not just commercial returns for private developers. The Aclimate and RRR cases exemplify this principle, while the NHS case shows the consequences of unclear value propositions.

Second, equitable data-sharing models must be adopted. Data collaborative approaches that distribute power and benefits fairly across stakeholders should be preferred over exclusive bilateral arrangements that concentrate advantages with large technology firms.

Third, human rights protection and data justice must be foundational. Every data sharing arrangement should undergo human rights impact assessments, with particular attention to effects on marginalized and vulnerable populations. The report emphasizes that data justice concerns are especially acute in LMICs, where regulatory capacity is limited and power asymmetries are pronounced.

Fourth, data minimization principles require that only the minimum necessary data be shared. The NHS case clearly violated this principle. Fifth, a clear legal basis must exist for all data sharing with third parties. Sixth, independent oversight institutions must be established or empowered to monitor data sharing arrangements. And finally, transparency and data subject participation are essential for building the public trust without which no data sharing program can succeed sustainably.

These principles are not merely theoretical. They represent practical guidance distilled from real-world successes and failures across diverse contexts, offering governments a roadmap for harnessing their data assets for AI development in ways that are responsible, equitable, and democratically accountable.