The geography of AI diffusion
A visual atlas of where generative AI is used, how adoption is changing, and the differences hidden inside national averages.
Concepts: AI Diffusion · Digital Divide
Executive summary.
Ten findings from Microsoft’s AI Diffusion research. Adoption means the estimated share of people aged 15–64 using generative AI during the reporting period; adoption changes are in percentage points (pp).
-
Global use is rising broadly.
Worldwide adoption reached 18.8% in Q2 2026, up from 17.8% in Q1 and 16.3% in H2 2025. In the supplied country dataset, 145 of 147 economies recorded an increase from Q1 to Q2.
-
The UAE and Singapore lead adoption.
The United Arab Emirates (73.3%) and Singapore (64.3%) have the highest Q2 2026 user shares. Ireland follows at 49.9%, then France at 49.6% and Norway at 49.4%.
-
South Korea is gaining fastest.
The largest Q1–Q2 2026 increases were South Korea (+3.5 pp), the UAE (+3.2 pp) and Japan (+2.2 pp), reaching 40.6%, 73.3% and 24.7% respectively. South Korea also leads growth since H1 2025, at +14.7 pp.
-
Some economies remain near 6% adoption.
Q2 2026 adoption is lowest in Cambodia (6.1%), followed by Afghanistan, Tajikistan and Turkmenistan (6.3% each). The weakest quarterly growth was in Papua New Guinea and Syria, the only declines in the CSV: −0.1 pp each, to 7.6% and 7.4%.
-
The North–South divide is widening.
In Q2 2026, adoption reached 28.8% in the Global North and 16.2% in the Global South. The gap grew from 10.6 pp in H2 2025 to 12.6 pp, as the North added more percentage points in both 2026 quarters.
-
Platform choice has a geopolitical pattern.
The H2 2025 report describes strong DeepSeek uptake in China, Russia, Iran, Cuba and Belarus, alongside parts of Africa. Microsoft connects its reach with restrictions on access to U.S. services, affordability and distribution partnerships. This is the report’s interpretation of observed patterns.
-
U.S. metropolitan adoption is twice rural adoption.
For Q1 2026, the U.S. report estimates average AI user shares of 32.9% in metropolitan counties, 21.8% in micropolitan counties and 16.2% in rural counties. These local estimates expose a substantial geographic divide within one country.
-
Young-adult communities stand out.
U.S. counties where more than 10% of residents are ages 18–24 average 28.6% AI user share, versus 20.3% elsewhere. The report associates younger populations and university communities with higher use; these county-level comparisons do not establish causation.
-
Access barriers extend beyond AI tools.
The Q1 2026 report compares internet access of 90.1% in the Global North with 65.7% in the Global South, using separate connectivity indicators. The H1 2025 research also identifies limited local-language support as a barrier, alongside electricity, connectivity and digital skills.
-
Learning has a larger role in Global South use.
Among the consumer Copilot conversations studied, self-directed learning’s share was 29.2% higher in the Global South than globally; schoolwork’s share was 32.3% higher. These are relative differences, not pp, and describe the sampled early adopters rather than all AI users.
Country comparisons follow the supplied CSV; report vintages and methods are documented below. Half-year and quarterly reporting windows differ. Adoption measures usage, not productivity or proficiency. Read the methods and limitations ↓
Adoption rises; the gap widens.
Share of people aged 15–64 who used a generative AI product during the reported period.
Growth starts from very different levels.
Every economy in the published dataset has a higher AI user share in Q2 2026 than in H1 2025. The scale of the increase varies substantially.
From H1 2025 to Q2 2026
H1 2025 on the horizontal axis; Q2 2026 on the vertical axis. Both axes 0–80%.
The ten largest gains
Change from H1 2025 to Q2 2026, in percentage points.
Scroll sideways for all periods. Scroll the table for every economy.
| Economy | H1 2025 | H2 2025 | Q1 2026 | Q2 2026 | Four periods · 0–80% | H1 → Q2 pp | Q1 → Q2 pp |
|---|
Source: AI_Diffusion_Q22026_Update.csv. The original CSV takes precedence in this view; report revisions are preserved separately.
National averages conceal local variation.
Explore the Q1 2026 estimates at three geographic scales. All counties appear in the distribution; every original record remains available below.
3,143 counties and county equivalents
Number of counties in each 2.5-percentage-point interval.
Scroll sideways for all columns.
Bars use a common 0–80% scale across all three views. The metro file contains 35 supplied metros, not a complete U.S. metro census. Identifier strings retain their original form.
The reports contain more than rankings.
Infrastructure, access, model performance, developer activity, and local context. Explore the numerical tables and labeled values recovered from each edition.
Original charts whose full point-level data is unavailable
Read every original page.
All 94 pages from the five reports are available here, including maps, unlabeled series, appendices, methods, and references.
The source reader loads original pages on demand. Original files are available as separate downloads and are not expanded in print view.
Download the original files · 12 files
Definitions, provenance, and limits.
What diffusion measures
The global reports define AI diffusion as the share of people aged 15–64 who used a generative AI product in the reported period. Aggregated Microsoft telemetry is adjusted for operating system and device share, internet penetration, and population. Usage does not measure capability, productivity, or economic impact.
How comparisons are made
Country changes subtract published rounded percentages. A rise from 10% to 12% is +2 percentage points. Periods are labeled categories; half-year and quarterly windows have different durations. No global or national user rate is calculated by averaging local percentages.
U.S. estimates have a distinct method
County figures are spatial small-area estimates, designed to address small samples, IP-geolocation noise, and infrastructure differences. State and metro aggregates use working-age population weights. The supplied CSVs contain no uncertainty intervals or population weights; none are fabricated here.
Design and completeness
Inspired by Edward Tufte’s principles: direct labels, common scales, quiet reference lines, small multiples, and data beside its explanation. All 3,376 CSV records and 3,817 rate observations are retained. Additional exact report data is transcribed; all 94 original pages cover the remaining graphical and narrative evidence.
Source inconsistencies and editorial decisions
Reproducibility and original-file fingerprints
This page preserves the 2025–2026 source collection accompanying Microsoft’s AI Diffusion research. Filters and derived changes run locally in your browser. Source downloads reproduce the original bytes. CSV exports from filtered views are derived convenience files. SHA-256 fingerprints identify each original file.
| Original file | SHA-256 |
|---|
Attribution and MIT Licence
Research and source data © 2026 Microsoft, produced by the Microsoft AI Economy Institute (AIEI). Source: microsoft/ai-diffusion-report. Repository materials are provided under the MIT Licence. The original notice is reproduced below. This visual presentation was prepared for in/out.