Governments Should Record When Policy Advice Relies on AI, Belfer Center Argues

From adoption to reliance

The analysis starts with the UAE. On September 2, the country's Cabinet launched the Cabinet AI Advisor, a system of 32 specialized AI advisers meant to support Cabinet work. According to the official announcement cited by the Belfer Center, it will analyze policies and legislation, assess their financial, economic, social and environmental effects, and present recommendations drawn from an approved set of inputs and sources.

The UAE has taken other steps since 2025. These include a regulatory intelligence ecosystem linking laws, judicial rulings and public services, a National Artificial Intelligence System to advise the Council of Ministers, a two-year target to move half of government sectors, services and operations to agentic AI, and an Artificial Intelligence and Data Authority reporting directly to the Cabinet.

The author argues that such announcements describe what AI is expected to do, but leave open a harder question. In the piece's words, "Adoption asks whether AI is being used. Reliance asks what its analysis is allowed to carry."

A wider pattern, with weak controls

The piece says the trend is not limited to the Gulf. Saudi Arabia's Public Investment Fund launched HUMAIN in 2025, and Qatar runs the GovAI program to speed up AI adoption across government entities.

The OECD's 2025 review of 200 government AI use cases found that nearly half sought to enhance decision-making, sense-making or forecasting. Some 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 AI used in at least one area of government in 35 of 36 OECD countries with available data. Only 13 reported using it to support policymaking. Only 14 countries reported requiring pre-deployment risk assessments, and 11 reported conducting post-deployment audits.

Existing frameworks, including a 2025 U.S. Office of Management and Budget memorandum, Australia's policy effective December 2025, Canada's Algorithmic Impact Assessment and the UK's transparency standard, are described as serious machinery. But the author says they tend to attach accountability to a system, use case, project or operational decision, not to the specific claim inside a piece of advice. "A record of use is not a record of reliance," the piece states.

The Deloitte example

The analysis points to a 2025 Deloitte report for Australia's Department of Employment and Workplace Relations, which reviewed welfare compliance. It 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. Released correspondence shows AI use for technical code analysis had been approved, but officials later sought clarification about tools used for other tasks, including summarizing legal proceedings and preparing citations. The author says this does not establish what the disputed material contributed to the findings, and illustrates the difference between reconstructing an AI contribution after errors surface and documenting its role at the time.

What the proposed record would contain

The author recommends that governments keep a record of reliance where a policy claim materially depends on AI-supported analysis, using existing assurance processes where 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
• the official who authorized reliance on it

Where AI gives advice directly, the record should capture material inputs and outputs, with a named official responsible for reviewing and authorizing reliance. The piece also proposes a removal test: if the AI contribution were removed, would the claim still stand in the same form, and would the recommendation change?

The burden should be proportionate. Cleaning language or low-consequence clerical work may need little assurance, while summarizing consultations, estimating fiscal effects, ranking alternatives or supporting Cabinet advice would need more. The author calls this "a small decision record, not another broad AI bureaucracy."

What it means for governments

The analysis ties the idea to sovereignty. Governments investing in sovereign AI will still depend on global chips, clouds, vendors, models and standards, so the author argues sovereignty cannot only mean owning every layer of the stack. In policy advice, it must also mean institutional control over judgment, including knowing when an AI system has shaped a policy claim.

For administrations rolling out AI advisers, the practical message is that formal human authority may remain intact while evidence shifts upstream into systems that sort, summarize and rank what officials see. Documenting reliance at the level of individual claims, the author suggests, would let departments show how a recommendation was formed before errors force them to reconstruct it.