How are governments using AI? New evidence from around the world

Artificial intelligence is already making its way into governments around the world. Public servants are using AI to summarize documents and search for information. Tax administrations are deploying it to detect fraud. Governments are developing chatbots to answer citizens’ questions, tools to diagnose disease, and systems to make public administration more efficient. But how widespread is AI use in government? And are governments building the institutions, skills, and data systems needed to use it effectively? New evidence from the AI and Data for Better Governance Survey, prepared for the World Development Report 2026 (https://www.worldbank.org/wdr2026/) provides a window into these questions. The survey collected information directly from governments in 60 economies across all income levels, including central digital agencies, ministries, and the officials responsible for management information systems. AI adoption is broad—but still shallow The first finding is just how widespread AI has already become in government. Most governments in our survey report using AI in some form. But much of this adoption remains relatively informal. Forty-four percent of governments’ internal AI use consists of individual public servants using AI for ad hoc tasks, such as searching for information, summarizing documents, or getting simple assistance. Giving individual employees access to generative AI can improve productivity. But it is different from integrating AI into the way an organization works. When AI becomes part of formal workflows, governments can establish common processes, monitor its performance, manage risks, and accumulate organizational knowledge. When adoption remains at the individual level, many of those benefits may remain with individual users rather than the institution. This pattern is particularly pronounced in lower-income economies, but even among upper-middle- and high-income economies, about one-third of governments report that their internal AI use remains ad hoc. Innovation is not confined to rich countries At the same time, our findings challenge the idea that meaningful AI innovation in government is limited to high-income economies. We find that income is not destiny for government AI adoption. For example, the Philippines uses predictive AI to identify taxpayers at risk of under declaring income. Thailand uses AI to detect tuberculosis and lung cancer from chest X-rays. Uganda has developed an AI assistant to answer human-resource queries from within its electronic document management system. The Democratic Republic of Congo uses an AI system to help correct state examinations. Burkina Faso has developed tools ranging from an assistant that helps citizens navigate the legal corpus to AI applications for copyright monitoring. AI is only as good as the ecosystem around it AI cannot compensate automatically for weak government systems. For AI to improve public administration, governments need reliable and data that can be easily shared across different systems, public servants who know how to use AI appropriately, organizational structures that support experimentation and learning, and governance arrangements that manage risks. Across all country income levels, data quality and availability are among the most commonly reported barriers to AI adoption. This is true in rich and poor economies alike. Without accurate, accessible, and interoperable government data, even sophisticated AI tools have limited material to work with. The constraints then diverge. In low-income economies, 58 percent of governments identify the lack of AI policies, guidelines, frameworks, or standards as one of their three biggest barriers to adoption. None of the high-income economies in our sample identify this as a major barrier. High-income economies have moved on to a different set of problems. Sixty-five percent identify privacy and ethical concerns as a major barrier, compared with 19 percent of low-income economies. Legacy IT systems, interoperability, and shortages of specialized talent also become more prominent constraints. This suggests there is no single AI-readiness checklist that applies equally everywhere. Governments need different complements depending on where they are starting. The emerging challenge of “shadow AI” Access to generative AI is expanding faster than the rules governing how it should be used. Nearly two-thirds of governments report providing public servants with licenses or official access to generative AI tools. Yet only 39 percent have formal ministry-wide guidelines governing their use, while another 38 percent are still developing them. This creates the possibility of what we call “shadow AI”: public servants using AI tools without adequate institutional oversight or guidance. The issue is not that governments should stop experimentation until every regulation is finalized. But experimentation needs to generate organizational learning. Governments need mechanisms to determine what data can be entered into AI systems, when human review is necessary, how outputs should be validated, and who is accountable when AI informs a government decision. From adopting AI to building better governments The question for governments, then, should simply be: How can AI help governments make better decisions and ultimately deliver better outcomes for citizens? AI will strengthen public administration only if it helps governments make decisions and take actions that are more informed, appropriate, and effective. AI adoption should be part of a broader effort to strengthen governments’ capacity to use evidence to make better decisions. Moreover, governments should not focus their investments exclusively on AI. Building the ability of public servants and institutions to work with data, conduct analysis, interpret evidence, and draw sound inferences can improve decision-making today while also creating the foundations for more effective AI adoption tomorrow. Our findings suggest four complementary priorities. First, invest in people and analytical capabilities. Public servants need the skills to work with data and evidence as well as enough AI literacy to understand what these tools can and cannot do, recognize their risks, and know when human judgment is essential. Second, strengthen the data foundations. AI does not make poor-quality or fragmented administrative data disappear. Investments in data quality, interoperability, secure data sharing, and digital public infrastructure remain fundamental. Third, build organizational capacity to turn experimentation into learning. Individual public servants experimenting with AI can generate valuable innovations, but governments need mechanisms to identify what works, incorporate successful applications into organizational processes, and scale them where appropriate. Fourth, place equal emphasis on governance and adoption. Guidelines, oversight arrangements, dedicated teams, budgets, and evaluation systems should evolve alongside experimentation. Governance should enable governments to learn from AI while ensuring that its use is responsible and appropriate. This blog draws on “How Is Government Using Artificial Intelligence?”, a background paper for the World Development Report 2026: The Promise of Artificial Intelligence, by Timothy Lundy, Manuel Ramos-Maqueda, Shyam Jayanti Patel, Daniel Rogger, Flavia Sacco Capurro, and Josefina Silva Fuentealba. The background paper is available online along with the broader set of World Development Report 2026 background papers available here.