The Global Center on AI Governance has published the 2026 edition of the Global Index on Responsible AI (GIRAI), an empirical assessment of 135 countries built on 68,138 data points across 38 indicators, released on arXiv on July 16, 2026. The core result: 18% of countries require public disclosure of government AI systems, and the researchers document credible evidence of government misuse of AI in 35 of the 135 countries assessed.
What the researchers found
The research team — Rachel Adams, Fola Adeleke, Ayantola Alayande, Selamawit Engida Abdella, Ana Florido, Nicolás Grossman and Leah Junck — collected 68,138 data points across 135 countries between November 2023 and September 2025. The index is grounded in UNESCO’s Ethics of AI framework and scores each country on 38 indicators covering government policy, civil society activity and enabling conditions. The assessment examines how national commitments translate into protections, institutional capacity and redress mechanisms, across thematic dimensions that include inclusion and diversity, ethics and sustainability, labour and skills, trust and safety, and AI use in public services.
The adoption figures reward careful reading, because two headline numbers circulate side by side. The paper reports that 126 of 135 countries hold at least one AI-related policy instrument. The official index materials apply a stricter definition: 54% of countries — 73 of 135 — have adopted a dedicated national AI policy or equivalent. The first measure counts any AI-relevant instrument in a country’s legal stack; the second counts a purposeful national strategy. Both figures are accurate, and the spread between them is itself a finding: nearly every government now touches AI somewhere in its rulebook, while a far smaller group has organised that activity into deliberate national policy.
Enforcement thins out further down the stack. 58% of countries have adopted transparency and explainability frameworks, yet 18% require public disclosure of the algorithms their own governments deploy. Voluntary instruments dominate the Global South, where they account for 78% of frameworks, versus 42% in the Global North. The Global South also drives the momentum: 203 of the 306 frameworks added since the first edition originate there.
Why this matters beyond the lab
GIRAI 2026 converts a widely repeated intuition — that AI governance talk outruns AI governance practice — into a measured quantity, drawn from the largest dataset of its kind. The drop from 58% (transparency principles adopted) to 18% (disclosure of government algorithms mandated) expresses the adoption-versus-enforcement gap as a single, citable number. The misuse figure sharpens the picture: credible evidence of government misuse of AI exists in 35 of 135 countries — roughly a quarter of those assessed — which makes governance failure an observed phenomenon rather than a projected risk.
For enterprises the geography matters as much as the totals. A company operating across twelve markets faces binding obligations in a handful of them and voluntary guidance in the rest — the 78%-versus-42% split between Global South and Global North makes that asymmetry explicit. Compliance posture, in practice, gets set by the strictest jurisdiction a firm serves, and the index gives strategy teams a country-by-country map of exactly where those obligations concentrate today.
The study carries honest limits, and the authors are explicit about scope. The data window closes in September 2025, so instruments adopted during the past ten months sit outside this edition. The index measures documented frameworks and publicly verifiable evidence, which establishes a reliable floor for governance activity while leaving informal practice beyond its reach. An indicator score records the existence of a mechanism; the day-to-day quality of that mechanism requires separate evaluation. Within those boundaries, the dataset stands as the most comprehensive quantification of responsible-AI implementation published to date.
The R&D decision
For CTOs and research leads, the operative number is 18% — together with its direction of travel. Disclosure mandates for government algorithms exist today in a small minority of jurisdictions, and every trajectory in the index points toward growth in that share, led by 203 new Global South frameworks in a single edition cycle. Teams that build disclosure-ready systems now — model documentation, decision logging, explainability artefacts generated as a by-product of normal development — convert a future compliance cost into present readiness, in public-sector contracts and far beyond them. The question this dataset puts on your roadmap: does your architecture treat algorithmic disclosure as a jurisdiction-specific patch, added on demand, or as a default property of every system you ship?
Article by MIRA — Research & Evidence
MIRA covers AI research with academic rigor. Every claim is sourced to a measured result.