Governments Need to Know When They are Relying on AI

On September 2, the Cabinet of the United Arab Emirates (UAE) launched the Cabinet AI Advisor, a system of 32 specialized artificial intelligence (AI) advisers intended to support Cabinet work. The official announcement says the system will analyze policies and legislation; assess their financial, economic, social, and environmental effects; link subjects to global best practice and national priorities; and present recommendations drawn from an approved set of inputs and sources. It will also work with ministers around the clock to follow up on Cabinet decisions and their implementation.

The announcement followed earlier steps. Since 2025, the UAE has approved a regulatory intelligence ecosystem linking laws, judicial rulings, executive procedures and public services; announced a National Artificial Intelligence System to serve as an adviser to the Council of Ministers, the Ministerial Development Council, and the boards of federal entities and government companies from January 2026; set a two-year target to move half of government sectors, services, and operations to agentic AI; and approved the establishment of an Artificial Intelligence and Data Authority reporting directly to the Cabinet.

These announcements describe what AI is expected to do inside government. The harder question is what follows after the output is produced: how much weight should officials give it, who may challenge it, and when should its analysis help establish a claim rather than merely assist in preparing policy advice? Adoption asks whether AI is being used. Reliance asks what its analysis is allowed to carry.

Across the Gulf, AI is being treated not only as a technology industry but also as part of public-sector capacity and national development strategies. Saudi Arabia’s Public Investment Fund launched HUMAIN in 2025  to invest across the AI value chain. In Qatar, the GovAI program aims to accelerate AI adoption across government entities, and a 2025 initiative by Qatar's Ministry of Communications and Information Technology and National Planning Council seeks to build AI-powered data analytics and strategic decision support for national planning.

The pattern is not solely regional. In its 2025 review of 200 government AI use cases, the Organisation for Economic Co-operation and Development (OECD) found that nearly half of them sought to enhance decision-making, sense-making, or forecasting. In a more detailed breakdown, 31 percent aimed to improve productivity in analytical tasks, while only 9 percent sought to automate mundane tasks. The OECD’s 2026 Digital Government Outlook found that AI was used in at least one area of government in 35 of 36 OECD countries with available data, but only 13 reported using it to support policymaking. Operational controls also lag: only 14 countries reported requiring pre-deployment risk assessments, while 11 reported conducting post-deployment audits.

For administrations working with large datasets and across several languages, AI can search legislation, translate material, synthesize consultations, compare jurisdictions, and test scenarios. That capacity can be valuable. Reliance concerns the role an institution permits AI-supported analysis to play in supporting a claim, estimate, ranking, risk assessment, budget case, or recommendation. The relevant unit is the relationship between the contribution and the claim, not the model in the abstract. A system may locate documents well and establish causation badly. It may compare international practice without showing whether that practice can be transferred. Asked to make the case for an option, it may produce a persuasive argument without testing the alternatives.

Debates over AI safety, economic change and geopolitical competition matter, but they do not settle how governments should govern their own use of AI. A September piece by members of the OECD Working Group on Agentic AI in Government asks what agentic systems may do and access on government’s behalf, under whose authority, with what audit trail, and how their actions can be stopped or reversed. A parallel question is needed for policy advice: what should AI-supported analysis be allowed to make persuasive before officials act? Formal authority may remain human while evidentiary authority shifts upstream, into the systems that sort, summarize, and rank what officials see.

The gap is visible even in established public-sector AI frameworks. They increasingly recognize that some AI outputs carry more institutional weight than others. A 2025 memorandum from the U.S. Office of Management and Budget treats AI as high-impact partly when its output becomes a principal basis for a decision or action affecting rights or safety and requires testing, impact assessment, monitoring, human oversight, and risk acceptance for such uses. Australia’s current policy for the responsible use of AI in government – effective December 2025 – sets requirements for accountable officials, internal registers, and impact assessments for in-scope AI use cases, subject to phased implementation. Canada’s Algorithmic Impact Assessment determines the  impact level of an automated decision system. The UK’s transparency standard requires records for in-scope algorithmic tools, although its mandatory scope concerns operational decisions about individuals, organizations, or groups rather than analytical models supporting broad policymaking. The UK’s AQuA Book separately provides guidance on analytical quality assurance, including AI and “black-box” models.

This is serious machinery, not cosmetic transparency. But it still tends to attach accountability to a system, use case, project, or operational decision. Policy reliance often happens at a more specific point: inside the advice itself. A department may know that a tool was registered, assessed, or disclosed without knowing which claim in a ministerial submission became more persuasive because of AI-supported analysis. The record may not show whether officials treated that output as background material, analytical support, or a recommendation; how they checked it; or who authorized reliance on it. A record of use is not a record of reliance.

The gap is not hypothetical. A 2025 report prepared by Deloitte for Australia’s Department of Employment and Workplace Relations to review welfare compliance contained incorrect references, including a fabricated Federal Court quotation. A revised version disclosed generative-AI use, and Deloitte agreed to repay the final installment of its  A$440,000 contract. The department maintained that the report’s substance and recommendations were unchanged. But correspondence released afterwards shows that, while AI use for technical code analysis had been approved, officials subsequently sought clarification about the tools used for other tasks, including summarizing legal proceedings and preparing citations. That may not have changed the recommendations, but it does not establish what the disputed material contributed to the findings. The episode illustrates the difference between reconstructing an AI contribution after errors are discovered and documenting its role when the analysis is prepared.

Governments should keep a record of reliance where a policy claim materially depends on AI-supported analysis, using existing assurance processes wherever possible. It need not log every prompt. It should identify the claim the AI-supported contribution helped establish, the role officials allowed it to play, the supporting evidence and sources, the checks performed, who could challenge or reject it, and the official who authorized reliance on it. Where AI gives advice directly rather than contributing to advice prepared by officials, the record should capture the material inputs and outputs, with a named official responsible for reviewing and authorizing reliance on the AI-supported advice. Where a claim may materially depend on AI-supported analysis, apply a removal test: if the AI-supported contribution were removed, would the claim still stand in the same form, and would the recommendation change? 

The burden should be proportionate. If AI cleans language or performs a low-consequence clerical task, little assurance may be needed. If it summarizes consultations, estimates fiscal effects, reviews legal options, ranks alternatives, or supports Cabinet advice, the level of assurance should reflect both how much the claim depends on the AI-supported contribution and what would follow if the claim were wrong. This is a small decision record, not another broad AI bureaucracy.

This is also a sovereignty question, but not only in the familiar sense. Governments investing in sovereign AI capability will still depend on global chips, clouds, vendors, models, and technical standards. Sovereignty therefore cannot only mean owning every layer of the stack. In policy advice, it must also mean institutional control over judgment: knowing when an AI system has shaped a policy claim, how that influence was tested, and on what basis officials accepted it.

The lesson from the Gulf, the OECD evidence, and the emerging agentic-AI debate is not that governments should slow down. It is that speed requires a clearer record of judgment. AI’s entry into policy advice should leave an intelligible account of what it carried: not merely that a system was used, but what the government allowed it to make persuasive.

Statements and views expressed in this commentary are solely those of the authors and do not imply endorsement by Harvard University, the Harvard Kennedy School, or the Belfer Center for Science and International Affairs.