Executive Summary

The definition of the metric is one line long. It measures, in Microsoft's words, "the share of people worldwide ages of 15 and 64 who have used a generative AI product during the reported period." The arithmetic behind that line is a product of three terms, and only the first of them is observed. The other two are estimated from outside data.

1.1The Limitations the Report Wrote Down First

A word on the character of this article, before anything else. Everything below is a recount, not an exposé. The README of the GitHub repository that holds the dataset has a section headed "Limitations," and the report put four lines there itself.

The same section closes this way: "No single metric fully captures global AI adoption. This dataset should be considered alongside complementary indicators where possible." The fourth paragraph of the September 21 company blog post opens on the same note: "No single metric is perfect, and this one is no exception." So what this article does is measure, against published material, how large the limitations those sentences point at actually are.

1.2The Method Is Not in This Report

The Q2 report runs eleven pages and has no methodology section. What it has instead is a "Citation and data availability" section on the last page that sends the reader to a different document. It says "This report is based on the following technical paper," names one arXiv paper, and adds that the same document is also available at aka.ms/AI_Diffusion_Technical_Report. Check whether the two addresses are different documents and they are not.

That paper is "Measuring AI Diffusion: A Population-Normalized Metric for Tracking Global AI Usage," posted on November 4, 2025, by five authors: Misra, Wang, McCullers, White, and Lavista Ferres. There is only a v1, and the data reaches to June 2025. The affiliation on it is the Microsoft AI for Good Lab, which is not the Microsoft AI Economy Institute that issued this quarter's report. The blog post carries the byline of the company's Chief Data Scientist, Juan Lavista Ferres. Everything from §1.3 onward comes out of that paper, and the party pointing readers to it is the report itself.

1.3One Observation Multiplied by Two Estimates

The equation in Section 2 of the paper is a product of three terms. The first is the share of Microsoft product telemetry users for whom a visit to an AI site was observed. The second is desktop device penetration per person aged 15 to 64, built by dividing Windows monthly active devices by Windows share of the PC and tablet market to estimate total devices, then dividing by population. The third is a scaling factor added to account for mobile usage, built from each country's ratio of mobile to desktop traffic. The source data for the second and third terms sits outside Microsoft. Operating system share and platform traffic ratios come from StatCounter; population comes from the World Bank's World Development Indicators, with CEIC filling in countries the World Bank misses.

The footnote attached to that last subtraction is worth reading. The paper adds that actual overlap between desktop and mobile AI usage is likely positively correlated, so the true overlap probably exceeds the estimate derived under independence, and that this "could result in a slight overestimation of the final combined AI User Share." That is a bias whose direction the people who built the metric named themselves.

The limitations the paper records about itself do not stop at that footnote. Section 4, the discussion, carries a broader statement. "Because our metric originates with Microsoft telemetry, it is inherently biased toward desktop platforms and the Microsoft user demographic. Although we apply rigorous adjustments and scaling factors, our results implicitly assume that user behavior in Microsoft products approximates that in other platforms, which may not always hold true." The remedies the paper proposes are all about widening the sources: integrating data from mobile app analytics providers such as Sensor Tower, or drawing on web traffic analytics from tools like Semrush and SimilarWeb. How far that assumption actually holds is something Section 5 looks at, by setting the metric beside survey values for the same countries.

1.4Who Counts as a User

The blog's definition is "people who have used a generative AI product during the reported period." That phrasing reads like a question about whether someone has ever tried it. The actual criterion, set out in Section 2.1 of the paper, is narrower.

Whether those 90 minutes are time spent in AI products or time spent across the Microsoft products that form the observation base is not specified in the paper. The reason it gives for introducing the threshold is low overall activity by the user, which makes the second reading the more natural one, though published documents cannot settle it. Either way one thing is settled. The numerator of this metric is not people who have tried AI but observed users who cleared a 90-minute monthly threshold.

The observation window also differs from the quarter on the label. Section 2.4 of the paper says that "we aggregate across data going back to late 2024 in all analyses unless otherwise indicated," and gives cross-country stability and the smoothing of short-term fluctuation as the reasons. A value labeled "Q2 2026" is therefore not a value for that quarter alone.

1.5What Gets Counted Is Public

Whether the list of measured tools is public was the first thing we wanted to know starting this article. It is. Appendix A of the paper, headed "List of Sites/Apps Included in Our Analysis," carries nineteen names exactly as written there.

Reading the list clears up a common misreading right away. Four Chinese tools are already on it: DeepSeek, Baidu's ERNIE Bot, Alibaba's Tongyi Qianwen, and Tencent's Xiaowei. Russian tools are represented by Yandex's Alice, Sber's GigaChat, and NanoSemantics, and Korea is represented by Naver's CLOVA X. So the claim that Chinese tools were left out wholesale is wrong. The question is which ones are missing, and Section 4 answers it by comparing lists.

The list is called "sites/apps," but the observation base is Windows desktop browsing, and mobile is estimated through the third term of Figure 1. An app that is mobile-only or mobile-first can have its name on the list and still never be directly observed.

1.6The Adjustment Inputs Are Written Differently in Different Documents

One discrepancy sits between the documents. The company blog states that the measure "is adjusted to reflect differences in operating system and device market share, internet penetration, and population," and the README sets out the same three items on separate lines. Yet none of the three terms in equation (1) of the paper corresponds to internet penetration. ITU internet data appears in the paper in exactly two places: a funnel figure showing technology penetration against national income, and a separate derived metric called AI User Share (Connected Population). Both sit outside the calculation of this metric.

On the most generous reading, the blog and the README may have loosely rendered the platform traffic ratio of the third term as "internet penetration." No public document confirms that. Only one version of the paper exists, and the Q2 report carries no methodology section of its own, so no document exists that resolves the discrepancy today. Nothing here is being called wrong. Two things hold and no more: the blog and the README name three items as adjusted, while the equation in the published methodology document multiplies two terms and then divides by population. And coverage after September 21 carried the phrasing about adjusting for device market and operating system, internet access, and population straight through.

Not one value we calculated is mixed into this section. It holds only what the report said, and the calculation was all pushed to the next section.

As of June 2026, 18.8% of the world's working-age population had used generative AI, about one percentage point above Q1. The blog describes the rise as broad: "almost every economy saw increased usage." The top of the leaderboard is unchanged from the previous round, with the United Arab Emirates at 73.3% and Singapore at 64.3%. The largest absolute increase belongs to South Korea, up 3.5 percentage points, and the largest climb in rank belongs to Saudi Arabia, from 30th to 25th. The largest relative increase belongs to Japan, up 2.2 percentage points from 22.5% in Q1, or roughly 10%.

The headline of this round was the gap. The blog's sentence reads: "the quarter brought a continued widening of the AI gap between the Global North and South, with usage now at 28.8% in the North and 16.2% in the South." The table below transcribes the figures the report has published across four rounds. The H1 2025 row comes from the Q1 report; the other three come from the Q2 report.

Table 1. Values as published by the report. The unit is percent of the population aged 15 to 64. The gap column is the difference between the two, and the report prints that column itself.

Drawn as lines, the same figures show two curves rising side by side while the space between them opens a little. Measured in percentage points, there is no arguing with that picture. Section 3 measures the same four points with different yardsticks.

Two other things came out with this round. One is an analysis of usage intent: in the Global South, self-directed learning and skill development take a 29.2 percent larger share of conversations than they do globally, and schoolwork and academic support a 32.3 percent larger share. The other is a section on open-weight models, which notes that the rising share of open-weight models in API token volume "could especially benefit the Global South by lowering access costs."

Now we open the public dataset. The repository's AI_Diffusion_Q22026_Update.csv is a table of four rounds of values for 147 economies, with not a single missing cell. The license is MIT. The decision to open that file is what made this article possible, and it deserves plain credit. The recount splits four ways: what reproduces, what was never counted, what wobbles inside the published precision, and what does not reproduce.

3.1What Reproduces

Country values and their changes come straight out of the public data. Recounting with our own script, the largest absolute gains from Q1 to Q2 are South Korea at 3.5 percentage points, the United Arab Emirates at 3.2, Japan at 2.2, Taiwan and Saudi Arabia at 2.0 each, and France at 1.8. Only two economies declined, Papua New Guinea and Syria, each by 0.1 percentage points. Of the 147, 71 did not move a single place in rank, and the steepest fall belongs to Papua New Guinea, from 129th to 138th. The unweighted mean of the 147 values rose from 19.07% in Q1 to 19.87% in Q2, and the median from 14.8% to 15.4%. China held at 67th, the United States held at 21st, and India slipped one place from 63rd to 64th.

We also checked against the leaderboard printed in the report. Matching the values for the top 30 in the PDF line by line against a recalculation from the public CSV leaves no row out of place. The full ranking down to 147th, though, is not in the PDF at all; it exists only in the CSV. One caveat has to be attached to the largest climb in rank, and that story belongs to §3.3.

3.2Thirty-Six Economies Were Never Observed on Their Own

Open the paper the report points readers to and something appears that the public CSV alone cannot reveal. Section 2.4 explains how the paper chose which countries to analyze individually.

In the full table of Appendix B, individual rows carry a dagger, and the legend under the table gives that dagger one line of meaning: "Region-imputed estimate due to insufficient telemetry coverage." Parsing Appendix B line by line gives 36 daggered rows. Elsewhere and independently, Section 3.3 of the paper refers to "the 111 economies in our study that we are able to estimate economy-specific estimates for (as opposed to regional averages for the economies with insufficient data)." Subtract 36 from 147 and you get 111. The paper's own number lands exactly.

A leaderboard of 147 economies is therefore not 147 measurements. One table holds 111 observed economies and 36 economies carrying an imputed regional average. The imputed clusters move as single bodies across all four rounds, because their values are identical for the whole period. Here are the six of them.

Table 2. The 36 daggered economies from Appendix B of the paper, grouped into clusters. Countries within a cluster carry the same value to one decimal place across all four rounds. Our verification script confirmed the match for the whole period.

The first row of that table is the most concrete single place in this article. Nigeria's published figure is not an observation of Nigeria but a West Africa average shared by twelve countries. By the World Bank's final 2023 figures, Nigeria has 126.4 million people aged 15 to 64, more than any other country in Africa. The values for Ethiopia (74.0 million) and Tanzania (36.1 million) are one number split eight ways. Add up the working-age population of all 36 imputed economies and it comes to roughly 600 million people, 11.5% of the world's population aged 15 to 64. Somalia is excluded for want of a final 2023 World Bank figure, so the real total is somewhat larger.

The tone here has to be exact. This is not concealment. The paper describes the method in Section 2.4 and marks it on every affected row of Appendix B. The trouble is that the place where it was disclosed and the place where it gets cited are not the same place. A reader arriving by citation sees the 147 rows of the CSV and the leaderboard in the report, and neither carries a dagger. Nothing in the dataset, the report, or the README has a column telling you which rows are imputed.

The paper presents this treatment as design rather than defect. It does so in Section 4, comparing its ranking against the World Bank's traffic-based one. Of that ranking, which places economies like Brunei, Suriname, and St. Kitts and Nevis at the top on ChatGPT site traffic per internet user, the paper reads the result as "likely reflecting outlier behavior from a small number of power users or data sensitivities inherent in measuring traffic within very small populations." Its own approach, it writes, "intentionally excludes markets with insufficient data coverage or combines with similar, neighboring countries to create a regional average, offering a more stable and comparable signal of AI adoption across countries and regions." The reasoning is sound. A stable signal beats a top of the table that swings on outliers from small states.

Suriname, though, sits inside the signal that was stabilized. It is one of the four rows in the Guianas cluster of Table 2, and those four countries can no longer be told apart. Stability was bought with resolution, and the only record of where that trade happened is the dagger in the appendix. What this section takes aim at, then, is not the judgment to use regional averages but how far the record of that judgment travels.

Imputation is also only one of two layers. Section 2.1 of the paper records a third layer separately. For countries where the vast majority of users opt out of sending telemetry, the sample may be too limited to reflect local trends, so the paper says it will "adjust AI usage estimates in markets with below-average opt-in rates by blending the observed country-level usage with the global average. The lower the opt-in rate, the more heavily we weight the global average." The paper describes those countries only as "particularly in parts of Europe" and names none of them. Because the treatment pulls values toward the global average, it compresses the differences between countries.

Putting the three layers on one page makes it plain how far the marking travels. No layer keeps its row-level mark in the final product, and only one, imputation, is marked as far as the paper's appendix. For blending, even the number of countries it applies to is unknowable.

3.3Inside the Rounding, the Ranking Wobbles

The dataset publishes values to one decimal place only. We measured how robust the ranking is at that precision on two bases: treating all 147 rows as one sample, and recounting with only the 111 observed economies left in. The figures from the two bases must not be read into each other. The first includes the imputed clusters; the second has them stripped out.

Table 3. The rounding band was taken as the published value ±0.05 percentage points. The 157 pairs and the 13-country cluster in the left column arise mostly from the imputation in Table 2. The right column is the distinguishability among actual observed values.

The difference between the two bases matters for a reason. Counted across all 147 rows, the largest tied cluster is 13 countries sharing 10.4%, and twelve of them are the West Africa cluster from Table 2. That is not a coincidence of rounding; it is the result of imputation. Countries that happened to round to the same value could not possibly move in lockstep to one decimal place across four rounds. Ties among observed economies, by contrast, are far scarcer. Two different mechanisms were mixed into one statistic, and any statement about the robustness of the ranking has to keep them apart.

Even so, one place remains where a genuine tie between observed economies shakes the ranking. The report's leaderboard prints Saudi Arabia 25th at 31.4% and Italy 26th at 31.4%. Neither carries a dagger in Appendix B, so both are observed economies whose values happen to round to the same place. Apply to the public CSV the simplest tie-break anyone would reach for, preserving the order the file itself uses, and sorting independently produces exactly the reverse order. Italy is the 24th row of the raw CSV and Saudi Arabia the 29th, so a stable sort puts Italy 25th and Saudi Arabia 26th.

That one place shakes one sentence. The blog wrote that Saudi Arabia "climbed five places from 30th to 25th, the largest rise in the rankings." Saudi Arabia stood 30th at 29.4% in Q1, so breaking the tie in its favor makes the climb five places and breaking it in Italy's favor makes it four. At four places it would equal South Korea (16th to 12th) and Thailand (87th to 83rd) in the same round, and the claim to the single largest rise would not stand. With the unrounded internal values Saudi Arabia may well come out ahead, and that is what reproduces the report's ordering. No one is alleging manipulation. This is a concrete instance of a limit that precision imposes: publish one decimal place and the rule a reproducer would sensibly pick can come out opposite to the publisher's actual assignment.

3.4What Does Not Reproduce

The three aggregates that went out in the headline cannot be calculated from published material: 18.8% for the world, 28.8% for the North, 16.2% for the South. The CSV holds rates for 147 economies and nothing else. No classification table saying which countries are Global North and which are South appears in the CSV, in the Q1 report, or in the Q2 report. The Q2 report offers only "we grouped AI usage by economy into the Global North and South, as above," and there is no list above either. The point of failure is not the weighted calculation but the roster the weights would be calculated over.

Even without the roster, its size can be worked backward. Treat the world value as a population-weighted average of North and South, and the North's population share is the world value minus the South, divided by the difference between North and South. Applying that to the four rounds gives 20.41%, 20.75%, 19.83%, and 20.63%, and allowing for one-decimal rounding bands, a single value somewhere between 19.69% and 20.83% explains all four rounds without contradiction. The report's Global North, in other words, has to be a grouping that holds about one fifth of the world's population aged 15 to 64.

So we checked three standard definitions against World Bank 2023 population. The World Bank's 86 high-income economies hold 17.19% of the world's population aged 15 to 64; the OECD's 38 members hold 17.22%; the 58 countries the United Nations counts as developed regions hold 15.64%. None of the three lands in the target band. Each falls short by four to five percentage points. We checked from the other direction too. Population-weighting the CSV under those same three definitions puts the South between 15.88% and 16.53%, close to the report's 16.2%, but puts the North between 30.67% and 32.41%, two to four percentage points above the report's 28.8%.

The two mismatches point the same way. The population share comes up short and the Northern average comes out high, so the report's Northern roster has to include more lower-usage countries than a standard high-income list does, and their population has to be added for the share to climb into the twenties. Beyond that, published material settles nothing. Rather than force a fourth definition into fitting, the honest conclusion at this point is not reproducible from published definitions. It is worth adding that the 147 economies in the CSV already cover 97.5% of the world's population aged 15 to 64. What is missing is not countries but the rule that divides them.

The same report contains a caption that shows what that one fifth weighs. The note attached to the usage-intent figure says the "North is 68.5% of traffic and South is 31.5%." One fifth of the population produces more than two thirds of the conversations under analysis.

Last, we measured the claim that the gap "continued to widen" with four different yardsticks. The conclusion first: percentage points run one way, and the other three change direction this round.

Table 4. All values calculated by us from figures the report published. The ratio and the log difference have overlapping rounding bands, so the fourth decimal place cannot be asserted. The possible band for the Q1 ratio runs from 1.7767 to 1.7948 and for Q2 from 1.7692 to 1.7864, and the two bands overlap.

The yardstick this article rests on is therefore relative growth. The North's 28.8% and the South's 16.2%, along with the 27.5% and 15.4% one round earlier, are all figures the report published, so the relative growth derived from them is untouched by the rounding argument. Q2 2026 is the first of the four rounds in which the South grew faster in relative terms. The headline for the same round is a gap that continued to widen, and that is also true. Both are true, and what gets measured decides the conclusion. For reference, treating the 147 countries as one unweighted sample, the Gini coefficient fell slightly from 0.3347 to 0.3333 and the ratio of the top decile to the bottom decile from 4.89 to 4.82. That calculation has a different unit from the report's population-weighted block comparison, so it should be read only as an appendix fact.

3.5The Report Was the One Asking for Reproducibility

Every recount above is work the report asked for itself. The citation notice in the Q2 report asks anyone using the metric to "cite the technical paper above and include a link to the dataset to ensure reproducibility and version transparency." Publishers who open their data and then ask that are rare. Two cases show how well the request holds up while the ranking circulates, and both come out of the publisher's own documents.

One is a rank. A September regional press release from Microsoft for Europe, the Middle East, and Africa says of Switzerland that it is "ranking 15th worldwide with an adoption rate of 39.1%, up from 37.8%." The two values are exactly right. Yet the Q2 report's own leaderboard prints Switzerland 14th at 39.1%, and our recalculation, sorting the public CSV in descending order, also gives 14th. On Q1 figures it is 14th as well. What causes the one-place difference cannot be established from published material, so we do not guess at it. That press release also carries no methodological caveat and no notice of the coming expansion of tools.

The other sits inside the paper. The body of Section 3.4 states that China's AI user share "has more than doubled from 8% to 20%, making it the world's largest AI market, with an estimated AI user base exceeding 195 million." Appendix B of the same paper prints China at 15.40%, and China's series in the public CSV runs 15.4, 16.3, 16.4, 17.5, never touching 20% in any round. The 20% in Section 3.4 is a three-month rolling average from Figure 6, so it is built differently, and the 195 million is consistent with 20% and China's working-age population. Which value is right is not something we settle. Only this goes on the record. One country's headline figure is written two ways inside the same document, and the published series matches only the lower of them. Set beside the README's warning that updates may revise past country-level estimates, the need for that request about version transparency comes into sharper relief.

There is a place where the report confirms the argument of the preceding sections itself. It is the passage where this round announces the revision coming in the next one. The fourth paragraph of the blog closes this way: "Over the past year, there has been a significant rise in new AI tools and models. For our next report, we will expand our measure to include usage of these new tools and expect this update to increase AI user share across almost all economies, with larger projected increases in China."

The body of the report says the same thing at greater length and adds one sentence. It is the sentence this whole article leans on.

The numbers going up next round is a change in the counting rather than a change in the world, and the publisher wrote that down first. A statement that plain is hard to get from the publisher of an indicator.

4.1The One Missing from the List Is China's Largest App

The report goes no further than "several widely used domestic tools" and never names them. But as §1.5 showed, the list of what is included is public, so what is missing comes out of a comparison rather than a guess. Matching the May 2026 monthly active user figures in the H1 AI application report that the Chinese mobile research firm QuestMobile released on July 14, 2026, against Appendix A of the paper looks like this.

Table 5. QuestMobile, H1 2026 AI Application Report (July 14, 2026), matched against Appendix A of the paper. QuestMobile measures only the mainland Chinese mobile internet, so overseas users of DeepSeek and Qwen are not in this table. And these are counts of app users while the Microsoft metric is a share of population, so the two cannot be compared as a multiple.

In the same survey, monthly active users across all Chinese AI-native apps came to 499 million, up 85.4% year over year, and Doubao alone accounts for 76.6% of that. It is larger than the second and third apps combined. So the one name missing from the list happens to be China's largest app. One aside belongs here: some English-language outlets rendered this figure as "3.8 billion." The original notation, 3.8亿, is 380 million.

Doubao is also a mobile-first app. As §1.5 recorded, mobile usage here is estimated, never observed. The first layer, absence from the list, and the second layer, mobile going unobserved, both land on the same app.

4.2Not a Limitation Revealed This Round but One Written Down All Along

The China coverage problem is not a new admission. Below Section 2.4 of the paper sits a separate box headed "Data Coverage Disclaimer," and it has read this way since November 2025: "In some countries—like Russia, Iran, and partly China—telemetry data is limited, so usage estimates may be incomplete. These gaps should be kept in mind when looking at country-level AI adoption." This round's sentence about domestic tools not currently being included restates that limitation, and what changed is that a revision was announced.

On the data side, that continuity looks like this. China's published value climbed across four rounds, 15.4, 16.3, 16.4, 17.5, while its rank did not move a single place, holding at 67th. Reading that stability as "AI diffusion in China has stalled" and reading it as "some of the tools widely used in China are not yet being counted" are completely different conclusions. And by the report's own notice, this country's value rises in the next round. Part of the reason the ranking looked stable was the scope of measurement.

This section does what the paper never did. The validation passage in Section 4 of the paper looks at how well its ranking agrees with other indices: a qualitative note that its top five countries also place inside the top twenty of the Oxford Insights Government AI Readiness Index, the Tortoise Global AI Index, and Stanford HAI's measures; "general alignment" with World Bank research; and an observation that Google Trends search interest supports its results. The only comparison for which a coefficient is reported is the rank correlation with income per capita, at 0.83. It has never been set level against a probability-sample survey. The paper cites Pew Research Center's finding that 23% of American adults have used ChatGPT, in Section 1, as a limitation of prior work, and does not place its own American value beside that survey. So the comparison table below fills a seat the paper left empty.

The rule for reading the table comes first. Every cell has a different denominator, method, and moment, so nothing has been done to them beyond setting them side by side. Multiples are stated only for pairs whose denominators are nearly the same, and where a conversion was needed, the conversion is spelled out below.

Table 6. The Microsoft column is a share of the population aged 15 to 64 and holds Q2 2026 values. Japan's approximate 24.7% is the 22.5% the blog gave for Q1 plus 2.2 percentage points. Values in the right-hand column each carry a different denominator, method, and moment, so they cannot be bound onto a single axis.

5.1Finding the Denominator Behind China's 42.8%

Usage rates from the China Internet Network Information Center are generally understood to take netizens as the denominator, so a reconversion looked necessary at first. Run the arithmetic, though, and this round is different. Dividing the announced 602 million users by China's total population of 1.4107 billion gives 42.68%, which matches the announced 42.8% within rounding error. Using the 1.125 billion netizens reported for the same period as the denominator gives 53.5%, which does not match. So this article reads the denominator for this round as the total population. We were not able to check the original footnote directly, and if that footnote specifies a different denominator, it should take precedence.

Lining the denominator up with the Microsoft metric takes one more step. Assuming all 602 million fall inside the 15-to-64 band and dividing by China's working-age population of 974.8 million gives roughly 61.8%. That figure is an upper bound. It rests on the ordinary pattern that generative AI use is rare among children and the elderly, and that assumption is not proven. Either way the direction of divergence from Microsoft's 17.5% is the same, and that direction matches the report's own notice. The Chinese center's count includes domestic tools in full, and Microsoft's current metric leaves some of them out.

5.2The Gap That Remains After Matching Denominators

The cleanest case in the table is the United States. Microsoft uses ages 15 to 64 and the survey uses adults 18 to 64, the closest-matched pair among the six countries, and still the values sit at 33.0% and 62%, nearly a factor of two apart. What separates the results is not the denominator but the difference between product telemetry and a self-reported survey. Japan runs the same way. From the nearby denominators of 15-to-64 and 15-to-69 come 24.7% and 58.8%, a factor of 2.4. For reference, one private Japanese survey came in at 54.7%, close to the ministry's figure, so two surveys agree in placing the number far above the telemetry.

South Korea shows the paradox running the other way. At 40.6% and 44.5% it is the closest pair among the six, yet the broader denominator belongs to the surveyor. With age 3 and older as the denominator, young children enter and the rate ought to fall, and instead the result is higher. The cause appears to be the measurement window. The survey asks cumulatively whether a respondent has ever tried a generative AI service, while Microsoft counts use during the reported period. It is a case of values inverting on a wider denominator once the measurement window differs.

India carries the largest gap, 18.5% against 63%, a factor of 3.4. Garimella's paper introduces its value as the first survey-based estimate of generative AI adoption in India, which means this comparison also shows that survey-based alternatives in India only recently came into existence.

5.3Two Cells We Could Not Fill

The Nigeria and Brazil cells were not forced full. Nothing was found that satisfies all three conditions: a nationally representative survey with published sample and methodology, or government statistics, and a usage rate against population rather than traffic. What turned up were values missing one condition each. "Over 80% of Nigerian professionals use AI regularly" rests on a sample of professionals, and "Brazil accounts for 5.3% of ChatGPT traffic" is a traffic share rather than a rate against population. Neither has grounds for conversion into the form of Table 6.

The blank itself is a finding. The first five countries in the table have alternative measures from government statistics or academic surveys, and the two that would be classified as Global South have none within the search. Measurement infrastructure is unevenly distributed by region, and in Nigeria's case another layer sits on top of that. Even the Microsoft-side value is not an observation of Nigeria.

What sits inside the denominator also differs by country. Pulling 2023 internet usage rates through the World Bank API gives South Korea 97.4%, the United States 93.5%, China 90.6%, Japan 85.0%, Brazil 84.2%, India 60.3%, and Nigeria 40.1%. Six tenths of Nigeria's population does not use the internet at all, while the denominator of this metric is the entire population aged 15 to 64, regardless of internet use. The ITU's 2025 summary points the same way: usage runs at 94% in high-income countries against 23% in low-income ones, with an African average of 36%.

Two articles on this blog have handled the same problem of measurement making the statistic, from different sources. One is a case where a work filter cut half the conversations out of an AI usage statistic, where the mechanism was how something gets classified as work, against denominators and coverage here. The other is the survey Gallup ran with Microsoft across thirty-seven countries, where that article watched trust and worry fall out of step within a single study, while this one watches how far a country's usage rate splits by method.

• A Work Filter Cut Half the Conversations Out of AI Use Statistics
• The people most worried about AI are not the ones who never use it

A procedure came out of this recount, not a verdict. The eight items below were raised out of the places where this article actually got caught while being written. Each carries a note on which passage of this article snagged on it. The list is not aimed only at the Microsoft metric; the same order works when pulling any external cross-country indicator into an internal document, and when building an internal indicator to send outside.

None of this means throwing out an indicator that fails the eight items. Do that and almost no usable cross-country indicator is left. It means citing the indicator with the items it failed written down alongside. Cited that way, the one line saying "our country ranks Nth" becomes a line saying "Nth on which list and against which denominator," and when the value moves next round, there is a way to judge what moved.

A model of that habit happens to sit inside the same report. In the open-weight section, the Q2 report cites another organization's public token statistics, and in citing them it attaches the limits of the denominator itself.

The habit is writing down the limits of the denominator within three lines of citing someone else's indicator. Asking the same of people who cite your own indicator is all it takes. And that request is already half written into the citation notice.

The problem this article has followed does not live only in national statistics. It has the same shape as a problem Pebblous meets again and again in data quality work. The three passages below are not about a product. They are about how the structure the earlier sections showed turns up again in practice.

7.1When Imputed Values Sit in the Same Column Unmarked

What §3.2 showed is the most familiar item in data quality practice. Original values and imputed values sit mixed in one column, and the final product has no mark telling you which row is which. Here the imputation was an average of neighboring countries, and twelve countries shared one value. Teams building dashboards meet this question every week. Which rows of this column are observed and which are filled in?

The difference lies not in the fact of imputation but in how far the mark travels. The paper described the method and flagged every affected line in an appendix, so the record on the data-processing side was kept. That record just did not make it to the CSV, the leaderboard, or the press release. Leaving lineage behind and making lineage travel all the way to the point of final consumption are separate jobs, and it is usually the second one that goes missing. Half of what Pebblous diagnoses through DataClinic is that second one.

7.2The Collection Path Sets the Boundary of the World

Product telemetry observes only what it can observe. Usage that goes to something other than a Microsoft product, or to a tool off the list, leaves no trace at all, and a rate cannot separate that from genuine absence. That is exactly the structure met in training data. The world the collection path defines becomes the boundary of the world a model knows.

This case goes a step further. Missing values are filled with a neighbor's average, thin samples are pulled toward the overall average, and the history of those treatments does not survive on the rows anyone finally reads. So when the report writes that next round's rise is "improved measurement," it is saying the same thing a quality practitioner says with the labels did not increase, the labeling scope did. That the publisher wrote this sentence in advance makes the report the rare case instead.

The same questions carry over when the object of diagnosis is an indicator rather than training data. The eight items of Section 6 each map onto one of the familiar axes of missingness, bias, and coverage. It is the point where the definition of AI-Ready Data widens from data that is good to feed a model to data that is good to make a judgment on.

7.3Recording the Caveats Instead of Discarding the Evidence

Client organizations set targets using external indicators. Sentences that run "our country ranks Nth in AI usage, therefore we will do the following" are common in planning documents. The evidence does not have to be thrown out. What it needs is a companion list: the denominator, the coverage list and its date, the provenance layer of the row, the published precision, and the revision history. The same list applies to an organization that builds an internal indicator and sends it outside.

What Pebblous can put forward here is one methodology. It is not the trust that comes from disputing someone else's numbers but the trust that comes from recounting them all the way through with published material alone and leaving the process behind. Which is why every derived value in this article carries a marker. Values back-calculated from a report table, values computed from public data, and values recounted with our own script are distinguished in the body text. The point is to hold ourselves to what we asked of someone else.

The evidence behind the body text runs along three strands. Items 1 through 6 are the Microsoft-side primary material that forms the spine of this article: the blog post, the full text of both reports, the public dataset and its README, the whole methodology paper, and a country press release, all read and transcribed directly. Items 7 through 12 are the policy and statistical material used to check the denominators and adjustment inputs, and items 13 onward are the academic and survey material behind the comparison table in Section 5 and the list comparison in Section 4. Item 4 in particular is the core evidence document for this article, and the party that told readers to go read it is item 2.

Primary Sources — Microsoft

• 1.Microsoft On the Issues (2026-09-21). "The continued state of global AI diffusion in 2026." The official blog post carrying the metric definition, the headline figures, and the notice of the coming expansion of tools. blogs.microsoft.com
• 2.Microsoft AI Economy Institute (2026-09). Global AI Diffusion Q2 2026 Trends and Insights. 11 pages. Leaderboard, Global North and South table, usage-intent analysis, open-weight models section, citation notice. It has no methodology section. github.com
• 3.Microsoft AI Economy Institute. Global AI Diffusion Q1 2026. The H1 2025 row of Table 1 and the North-South infrastructure comparison figures come from this round.
• 4.Misra, A., Wang, J., McCullers, S., White, K., & Lavista Ferres, J. (2025). Measuring AI Diffusion: A Population-Normalized Metric for Tracking Global AI Usage. arXiv:2511.02781 (v1, 2025-11-04). Equation (1), the 90-minute threshold, the 2-million threshold and regional grouping, opt-in blending, the Appendix A tool list, the row-level daggers in Appendix B. The core evidence document for this article. arxiv.org
• 5.The README and AI_Diffusion_Q22026_Update.csv (147 rows × 4 rounds, MIT license) in the microsoft/ai-diffusion-report repository. Every recalculation in Section 3 came out of this file. github.com
• 6.Microsoft News Source EMEA (2026-09). Switzerland press release. "ranking 15th worldwide with an adoption rate of 39.1%, up from 37.8%." The verbatim quotation used for the rank comparison in §3.5.

Policy and Statistics

• 7.World Bank, World Development Indicators. Indicator codes SP.POP.1564.TO.ZS, SP.POP.TOTL, IT.NET.USER.ZS, final 2023 figures. Used for the population-share check, the population sum of the 36 imputed economies, and the internet usage table.
• 8.ITU, Measuring digital development: Facts and Figures (2024/2025). 94% in high-income countries against 23% in low-income ones, with an African average of 36%.
• 9.China Internet Network Information Center (CNNIC), Statistical Report on China's Internet Development, 57th edition (as of December 2025). 602 million generative AI users, a 42.8% usage rate. The denominator judgment in §5.1 was made from the arithmetic of the announced figures; the original footnote was not checked directly.
• 10.Ministry of Science and ICT and the National Information Society Agency (NIA), 2025 Survey on Internet Use (South Korea). 50,750 people in 22,671 households, age 3 and older, 44.5% having used a generative AI service.
• 11.Japan Ministry of Internal Affairs and Communications (MIC), four-country comparative survey in the 2026 White Paper on Information and Communications. Self-reported, ages 15 to 69, Japan at 58.8%, fielded January to February 2026.
• 12.StatCounter, Desktop & Tablet OS Market Share and Desktop vs Mobile. The origin of the adjustment inputs used by the second and third terms of equation (1) in the paper.

Academic and Survey

• 13.Bick, A., Blandin, A., & Deming, D. The Rapid Adoption of Generative AI. National survey in the Real-Time Population Survey series. Against adults aged 18 to 64, 62% in May 2026 and 45% in August 2024.
• 14.Garimella, K. (2026). Survey of generative AI adoption in India. Online panel run alongside in-person fieldwork in rural Uttar Pradesh, reweighted to national age, gender, and urbanization figures. Non-work use 63%, work use 35%.
• 15.QuestMobile, H1 2026 AI Application Report (2026-07-14). As of May 2026, Doubao at 382 million, Qwen at 167 million, DeepSeek at 130 million, and all Chinese AI-native apps at 499 million. Confirmed through TechNode's coverage of 2026-07-14.
• 16.Google (2026-07). AI & Economy ATLAS v1.0: Mapping Gemini Usage in the Economy. The document the Q2 report cites as a parallel indicator. ai.google

Adjacent Pebblous Articles

• 17.Pebblous, "A Work Filter Cut Half the Conversations Out of AI Use Statistics." report/ai-use-measurement-independent-corpus-2026-08
• 18.Pebblous, "The people most worried about AI are not the ones who never use it." report/ai-daily-use-worry-gallup-2026-09