2026 Edition · 195 governments · 69 indicators Under development
National AI capability, measured on a comparable basis.
The Global AI Rankings™ assess 195 governments across six equally weighted pillars, synthesized from publicly available data and normalized to a single 0–100 scale so that capability can be read across borders rather than country by country. The composite rule is published in full.
Synthesized from public data · 11 August 2026 Full rankings →
Overview Under development
Assessing capability and control together.
Published strategy is now near-universal among assessed governments and no longer distinguishes between them. The Global AI Rankings therefore score what a state has established: institutions, compute and data infrastructure, workforce capability, and enforceable safeguards. Capability and control are weighted equally.
Since the first edition in 2025, the Global AI Rankings have addressed a single question: what does a government require to deploy artificial intelligence effectively and accountably, and how far has each one progressed? The scope is deliberately narrow. National AI industries, frontier model capability and private-sector adoption are assessed only where the state’s own capacity depends on them.
Policy capacity is now close to universal: 91% of assessed governments record a policy-capacity score above zero. Substantially fewer have established the machinery that converts a stated position into a governed deployment — a named accountable owner, a recurring budget line, a procurement route capable of acquiring AI systems safely, or a redress mechanism for individuals affected by automated decisions. The Rankings measure the distance between those two positions.
The 2026 edition covers 195 governments, built on the 69 indicators the source framework assesses across 14 dimensions. Where a government has not supplied evidence, the indicator is recorded as absent rather than estimated, and the omission is disclosed. Published gaps are preferred to inferred values.
Indicator-level data is published concurrently with the report. Governments, organizations and researchers may dispute a score through the corrections process. Every accepted correction is logged publicly with its date and its effect on the published result.
Assessment framework Under development
Six pillars, weighted equally.
Equal weighting is a deliberate choice rather than a default. A government cannot offset absent safeguards with abundant compute capacity, so nothing here is allowed to buy its way past anything else.
Policy capacity
Whether a government can design and fund AI policy against a stated national vision — including whether commitment is backed by assigned resources rather than published intent, and whether it engages internationally on a technology that does not respect borders.
AI infrastructure
Compute capacity and the enabling technical infrastructure that government and others can rely on over the long term, together with data that is sufficient, high quality, and able to move securely within and across borders.
Governance
Whether governance principles exist, are promoted, and shape how AI is developed and adopted — protecting rights and societal interests through the government's own practices, and giving those subject to AI decisions an adaptive route to compliance and redress.
Public-sector adoption
Whether digital policy actually encourages testing AI against public-sector problems, and whether what proves effective is then scaled — evidenced in e-government delivery rather than in pilots.
Development and diffusion
Whether a domestic AI sector is mature enough to serve local demand and sustain long-term innovation, with the human capital for AI to diffuse to industry, researchers and civil society beyond the specialists.
Resilience
Whether the government is managing the social, economic and environmental strain of widespread adoption, and actively monitoring the safety and security risks that grow as development and diffusion progress.
Findings
Six findings from the 2026 edition.
Findings are supported by indicator-level data released with the report.
The counts in this section are illustrative drafting copy, not results from the assessed register. They are written to show the shape a finding takes and must not be quoted. The findings that are computed from the published register are set out in the 2026 report preview.
Finding 01
The gap is no longer strategy. It is delivery.
Most assessed governments now have a published national AI strategy or one in development. On that measure the world is close to saturation, and the indicator has lost most of its power to discriminate between governments.
What separates the top quartile from the rest is unglamorous: a named senior official who can be held to account, money that recurs rather than arrives once, and a procurement route that lets a department buy a model-backed service without inventing the contract from scratch each time. Only 38 governments satisfy all three.
This is the clearest structural finding of the 2026 edition. Readiness has moved from a question of whether a government has decided to act to a question of whether it has built anything that can act.
Finding 02
Compute access is concentrating faster than capability.
Public-sector compute allocations grew in 61 assessed governments this year, but three-quarters of that new capacity sits in twelve countries. For most governments, the practical constraint is no longer whether AI systems are technically available — commercial APIs have largely solved that — but whether they can run sensitive workloads under their own legal control.
Several governments have responded by buying sovereignty at the infrastructure layer while remaining dependent at the model layer. The Rankings record both positions separately, because they fail in different ways.
Finding 03
Oversight is being written faster than it is being staffed.
Forty-four governments introduced or expanded AI oversight obligations during the assessment window. Of those, nineteen created a supervisory function without a corresponding increase in specialist headcount, and eleven assigned the duty to a regulator that had not previously employed anyone with machine-learning expertise.
Rules without inspection capacity score poorly in this edition, and deliberately so. The governance pillar weights enforcement evidence — audits conducted, systems withdrawn, penalties issued — above the existence of a legal instrument.
The governments that moved up most in this pillar were not the ones that legislated most recently. They were the ones that published what their oversight bodies actually did.
Finding 04
Public-sector deployment has outpaced public-sector evaluation.
Assessed governments reported 2,140 AI systems in production use, roughly double the 2024 figure. Documented post-deployment evaluation exists for fewer than one in five. Where evaluation does exist it is heavily concentrated in health and revenue administration — the two domains where an error has an immediate, individually attributable cost.
Redress lags further still. Sixty-two governments now require some form of human review of automated decisions; twenty-three publish how often that review overturns the original outcome. Without the second number, the first is a policy rather than a safeguard.
Finding 05
Mid-tier governments are gaining on capability, not on money.
The largest score improvements in this edition did not come from the highest-spending governments. Eleven upper-middle-income states moved up five or more places, almost entirely through the human capital and institutional pillars: shared service teams, secondment routes into departments, and standing arrangements with domestic universities.
These are cheap interventions with slow payback, which is precisely why they are under-represented in strategies written for a single electoral cycle. The Rankings reward them because they are the interventions that survive a change of government.
Finding 06
Divergence between blocs is now measurable.
For the first time, the variance in governance-pillar scores between regional blocs exceeds the variance within them. Governments are converging on their neighbors and diverging from the rest, a pattern that did not appear in editions before 2024.
The Rankings report this without prescribing a remedy. Whether regulatory divergence is a cost or a legitimate expression of differing public preferences is not a question an index can settle, and the pillars are not weighted to imply an answer.
An index is only useful if it is willing to publish a score its own funders dislike. That constraint shapes every methodological decision we make.
Regional analysis
Where the movement is.
Regional median 81.9 · 2 governments assessed
The region holds the highest overall scores in the register, driven by compute availability and a research base departments can draw on directly. Institutional scores are more uneven: federal structures make a single accountable owner harder to identify, and sub-national deployment often runs ahead of central oversight.
Regional median 69.7 · 24 governments assessed
Western Europe leads on governance and safeguards by a clear margin, with the highest concentration of governments publishing enforcement outcomes rather than obligations alone. The region loses ground on development and diffusion, where sovereign compute commitments remain largely prospective.
Regional median 62.4 · 17 governments assessed
The widest internal spread of any region: governments in the global top ten sit alongside governments below the global median. Where central coordination exists it tends to be well resourced and durable across administrations, which is what separates the top of the region from the rest of it.
Regional median 43.7 · 19 governments assessed
Public-sector adoption is the region's strongest pillar, largely through dedicated AI bodies with standing budgets. Development and diffusion remains the binding constraint, and several governments are addressing it through recruitment rather than domestic pipeline development.
Regional median 27.1 · 49 governments assessed
The largest region in the register by count, and the lowest by median score. Policy capacity is the pillar on which its governments come closest to the global distribution; infrastructure indicators continue to cap achievable scores irrespective of policy quality.
Reports and data
Everything the Rankings publish, in one place.
Reports and datasets are free and download directly — no form, no email address. If you would like to be told when a correction changes a score you may have cited, the form at the bottom of this page is the way to arrange it.
Pillar dataset (CC BY-SA 4.0)
Not yet publishedOnly the current edition is published here. Where a correction changes a figure, the revision is recorded in the corrections log and a new version of the file is issued — superseded figures are logged rather than left in circulation as an archived download.
Correction notices
Optional. Nothing above is withheld pending this, and the downloads work without it.
Discuss the 2026 findings with the research team.
Methodology questions, indicator-level data requests, briefings for policy and executive audiences, and media inquiries are handled directly by the assessment team.