Research Brief · Xiaohu Explains

OpenAI publishes country-level ChatGPT usage data: nearly half of work messages ask AI to get things done

Per-capita rankings for 144 countries, three-year growth multipliers across six continents, and users 35 and older gaining share in over 90% of countries — most of these numbers only appear in the charts, not in the report's text.
The 60-Second Take
  • OpenAI has opened up its data on how different countries use ChatGPT: rankings for 144 countries, growth curves by continent, and share shifts by age group. The CSV files are directly downloadable.
  • In work contexts, 45% of messages ask AI to produce or execute something. Outside work, that number drops to 22%.
  • Several key figures only appear in the charts — the report's text never mentions them.
⚑ Data published and self-reported by OpenAI's economic research team. The "over 1 billion users" figure is their own estimate, unaudited by any third party. Per their public methodology, all statistics are based on a monthly sample of 300,000 messages, with differential privacy noise added — so every percentage below carries some margin of error and shouldn't be treated as exact. Charts come from the original report; this site cross-checked every number against them. Values not labeled on the line charts were read by this site from high-resolution images and are marked "approx."
The Release

OpenAI Publishes Country-Level ChatGPT Usage Data for the First Time

OpenAI released a data report yesterday that breaks down how people around the world actually use ChatGPT, country by country — the first time it's done so. Previously it only published global totals; now it's laid out over a hundred countries, and the CSV files are ready to download.

For anyone wondering how far AI has actually penetrated, this is the only official dataset you can slice by country right now. The headline finding boils down to one line: people are shifting from "asking AI" to "delegating to AI."

The data comes from OpenAI's economic research team and lives on a platform called OpenAI Signals, where the methodology PDF is also public.

One Definitional Note First

This dataset only covers messages sent from personal accounts at the Free, Go, Plus, and Pro tiers. Business and organization accounts are completely excluded.

So every number below speaks to "how individuals use it." Even when "work context" appears, it refers to messages sent from personal accounts that were classified as work-related — not how company-purchased ChatGPT is used. Without flagging this upfront, every figure that follows would be misread.

Core Finding

In work contexts, nearly half of messages ask AI to get things done

OpenAI sorts all messages into three categories:

Doing
Asking it to execute a task or produce something — editing a draft, writing code, running an analysis.
Asking
Looking up information or requesting an explanation. You ask, it answers.
Expressing
Self-expression — talking through feelings, writing something personal.

Set work and non-work contexts side by side, and the gap jumps out:

SHARE OF MESSAGES (GLOBAL) Work Doing 45% Asking 32% Expressing 22% Non-work Doing 22% Asking 45% Expressing 33% Doing in work contexts is more than double its non-work share
This site's Chinese-language comparison, drawn from the original report's chart values. The original was an English stacked bar chart with identical numbers.

Here's the original chart OpenAI provided, with the three colors mapping to the categories above:

OpenAI original chart: stacked bar chart of message category shares in work versus non-work contexts
Original report chart, "For work, individuals messages are more about doing." Left bar is non-work, right bar is work, orange is "doing." The chart includes a Country dropdown to view individual nations. Source: OpenAI.

The flip side:in non-work contexts, Asking remains the largest category at 45%. After hours, people are still exploring and querying more than they're delegating.

One easily missed detail: "non-work" is a broad bucket. The dataset actually splits things into work, study, and other — but on this chart, the report only uses the work / non-work binary, so students using ChatGPT for homework are counted on the non-work side.

We covered the product this report references
OpenAI Launches ChatGPT Work, Powered by GPT-5.6, Auto-Producing Docs, Spreadsheets, and Slides Across Tools
OpenAI ties the rise in "doing" within work contexts directly to the launch of ChatGPT Work.
Global Distribution

Regions that started late are growing fastest — Africa up 24× in three years

One chart plots curves by continent, with the Y-axis showing how many times weekly active users have multiplied relative to July 2023, running through June 2026.

OpenAI original chart: line chart of three-year weekly active user growth multipliers across six continents
Original chart, "ChatGPT adoption is growing fastest in regions with lower initial adoption." Six lines, top to bottom: Africa, Asia, South America, Europe, North America, Oceania. Source: OpenAI.
Growth multiplier over three years (read by this site from the chart above)
Africa~24×
Asia~16×
South America~11×
Europe~11×
North America~9.3×
Oceania~6.7×
Don't Read This Chart Backward

A higher multiple doesn't mean more usage. North America's lowest multiplier (~9.3×) is precisely because it started with the largest base — ChatGPT's home turf. Africa's 20+ fold growth reflects a near-zero starting point.

This chart shows the gap narrowing, not the rankings reversing.

Quarterly Changes

Peru rose 17 spots in one quarter; Italy fell 11

OpenAI also ranked 144 countries by "per-capita message volume," then charted who rose and who fell over the quarter. Blue means a ranking gain, brown a drop, and gray hatching means no operations or insufficient data.

World map: change in per-capita message rankings for 144 countries from Q1 to Q2 2026
Original chart, "Latin America and Oceania led Q2 adoption-rank gains." Color scale from −20 to +20, with Peru, Uruguay, Costa Rica, Australia, New Zealand, Italy, and Austria labeled. Source: OpenAI.
Biggest gainers
Peru+17
Uruguay+16
Costa Rica+15
Australia+9
New Zealand+9
Biggest drop-offs
Italy−11
Austria−9

These two only appear in the map's labels — the report's text never mentions them. The text says "Latin America, Oceania, and Africa grew faster than the rest of the world, and adoption in North America and Europe continued to rise," which matches the patch of brown across Europe on the map.

The report offers no explanation for Italy and Austria's ranking drops, and this site couldn't find one either. Rankings are relative — if others grow faster, you get pushed down.

Usage Shifts

People making images and videos with AI nearly quadrupled in two years

Trace the share of each use case from August 2024 to June 2026, and the fastest riser is multimedia — generating, analyzing, and retrieving images and video.

OpenAI original chart: two-year trend in message share by use case, with the multimedia line highlighted
Original chart, "Multimedia use grew fastest relative to other use cases." The blue-highlighted line is multimedia; the gray lines below it are, top to bottom: practical guidance, writing, finding information, self-expression, other, and technical help. Source: OpenAI.
Multimedia · Fastest growth
~2% → 7.8%
Nearly fourfold in two years. OpenAI attributes this to ChatGPT Images 2.0, released in April 2026.
But still fourth in absolute share
Top three unchanged
Practical guidance ~31–33%, writing ~22.5%, finding information ~18.5% — all still ahead of multimedia.

The same chart shows one line heading down: self-expression has slipped from ~13.5% to ~9%. Writing is also slowly declining, from ~26% to ~22.5%.

What's Actually Inside These Six Categories

The category names on the chart are vague. "Practical guidance" taking up a third of messages isn't self-explanatory. The methodology document breaks down which fine-grained labels roll up into each category:

Practical guidance · ~31–33% (largest)
How-to advice, tutoring and instruction, creative ideas, health/fitness/beauty and self-care.
Writing · ~22.5%
Editing or critiquing text you provide, personal writing and communication, translation, argumentation or summarization, fiction writing.
Finding information · ~18.5%
Looking up specific facts, researching products to buy, cooking and recipes.
Self-expression · ~9%
Greetings and small talk, relationships and personal reflection, games and role-play.
Multimedia · 7.8%
Generating images, analyzing images, generating or retrieving other media.
Technical help
Math, data analysis, programming.
A Hidden Counterintuitive Detail

Writing code lands under "technical help," not "writing." So the "writing at 22.5%" figure has nothing to do with programmers — it covers editing text, translating, fiction. Programming sits lumped in with math and data analysis.

Multimedia also shows a clear spike in April 2025, hitting ~9.5% before settling back near 7% and only later climbing to 7.8%. The report offers no explanation for that spike, and this site won't speculate either.

Latin America leads; one in ten Brazilian messages is multimedia

Break multimedia down by country, and the regional pattern is obvious:

World map: multimedia message share across 126 countries, color scale 0% to 15%
Original chart, "Latin American countries had higher-than-average multimedia use." Covers 126 countries, scale 0%–15%, with Brazil, Colombia, Mexico, and Japan labeled. Source: OpenAI.
11.1%
Brazil — highest globally
10.1%
Colombia
4.7%
Japan — labeled low for contrast

Both Brazil and Colombia cross the "one in ten messages is multimedia" line, with Mexico landing right at 10%.

User Demographics

People over 35 are catching up to younger users

Start with the global picture. Two lines track each age group's share of messages:

OpenAI original chart: message share curves for the 18-34 and 35+ age groups
Original chart, "People over 35 increasingly used ChatGPT over the past year." Orange is 18–34, blue is 35+. ⚠️ Per the methodology document, the whole dataset only includes accounts that declared an age of 18 or older; under-18 and age-unspecified accounts are excluded. Source: OpenAI.

The 18–34 share peaked around 75.5% in late 2024, then drifted down to roughly 66.5%. The 35+ group moved in the mirror direction, climbing from a low of about 24.8% to around 34%. The gap is closing.

The report collapses age into two buckets, but the underlying data actually uses six: 18–24, 25–34, 35–44, 45–54, 55–64, and 65+. The "35+" line on the chart is the sum of the last four buckets — downloading the CSV reveals the finer distribution.

By country: Czechia up 12.6 percentage points in a year

The last chart plots the change in the 35+ share for 111 countries, using Q2 2025 as the baseline.

OpenAI original chart: change in 35+ user message share across 111 countries relative to Q2 2025
Original chart, "Users aged 35+ gained share in over 90% of countries." Gray lines are all 111 countries, the dashed line is the median, and the colored lines are Czechia, France, South Korea, US, and Singapore. Source: OpenAI.
Change in 35+ message share over one year
Czechia+12.6
France+10.1
South Korea+8.0
United States+7.8
Median of 111 countries+5.1
Singapore+1.2

Units are percentage points. Outside the median and the labeled countries, the remaining 100+ lines are gray and unlabeled on the chart.

Percentage Points vs. Percent — Not the Same Thing

Going from 30% to 35% is a gain of "5 percentage points." Saying "up 5%" would only take you to 31.5%. The report text says the 35+ share was "5% higher," while the chart labels the median as "+5.1 percentage points" — a meaningful difference. This site follows the chart's percentage-point framing.

The chart title is also more specific than the text: the title says "over 90% of countries," while the text only says "almost every country."

Regional differences are visible on the chart too: roughly three-quarters of European countries gained more than the average. Southeast Asia runs the other way — six of eight countries are up, but only modestly; Singapore sits at just 1.2 percentage points.

We covered this team's previous study
OpenAI Sifted Through 800K ChatGPT Work Messages: In Five of Eight Job Roles, Over Half of Professional Queries Were Doing Someone Else's Job
That study dug inside work messages — how much of what people do with ChatGPT falls outside their own job description.
Boundaries

This dataset only counts personal accounts — it says nothing about enterprise usage

Before drawing conclusions from these numbers, a few caveats need to be on the table. They're not in the report itself; they're in the accompanying methodology document.

Counted Messages from personal accounts, 300K sampled monthly Free · Go · Plus · Pro Only accounts declaring age 18+ Not counted Business and organization accounts Codex Users who deleted accounts or opted out of training So "work context" = work-classified messages from personal accounts, not how company-purchased ChatGPT is used
Compiled by this site from the OpenAI Signals methodology document (README v2.0).
OpenAI's Own Admission

The methodology document includes this line: the dataset excludes enterprise message classifications, so it likely undercounts commercial usage.

The "45% doing in work contexts" figure measures how individuals use their own accounts at work. The company-purchased side is entirely invisible here — and per OpenAI, real business usage is higher than what's shown.

Four More Details That Affect How You Read the Numbers

It's a sample, not the full corpus
From July 2024 through June 2026, 300,000 messages were sampled each month for classification. So any precision to one decimal place is still an estimate at heart.
Numbers have noise added
To protect privacy, differential privacy noise is injected into the statistics before publication, so no specific message can be reverse-engineered from the public figures. The trade-off is that every percentage carries a bit of error.
Classification is done by models, not humans
Labels like "doing vs. asking" and "work vs. non-work" are assigned automatically by large-model classifiers. Researchers never read raw user messages; content-based analysis runs on de-identified classification results.
Per-capita figures use 2023 population
Country rankings normalize by World Bank 2023 population estimates — a base that's three years old.
The Finer the Slice, the Shakier the Number

Every series comes from the same 300K-message monthly sample. Cut it finer and fewer messages land in each cell, inflating variance. The document explicitly flags granular cuts like "share of a specific topic in a specific state."

By that logic, big-picture numbers like the global 7.8% are the most stable; single-country, single-category figures like Brazil's 11.1% or Japan's 4.7% should be taken loosely, not compared at the decimal level.

Figures marked "approx." in this article were read from high-resolution charts where the report provided no data table.

Things in the Dataset the Report Didn't Use

The report only uses part of the data. The full package of 25 CSV files includes dimensions that never made it into a chart:

Gender
Inferred from first names, not self-reported — anonymized names are matched against external name-to-gender sources, and only names with >95% confidence in one direction count. Available by month, country, and topic.
US state-level
2025 per-capita rankings and topic distributions by state, normalized by US Census Bureau 2023 estimates.
Occupational activity classification
US data is coded against O*NET "intermediate work activities," showing which types of professional actions work messages map to.

Back to what this data is actually worth. Any single number in isolation is of limited value. What matters is that "how far has AI penetrated" can now be answered by country for the first time: previously there was only one global total; now you can see Peru climb 17 spots in a quarter, one in ten Brazilian messages generating images, Czechia's middle-aged users up 12.6 percentage points in a year. Three threads run through it all: AI is already doing the work, late-starting regions are catching up fast, and the users are no longer just young people.

🧰 Quick Card · OpenAI Signals (data platform)
PriceFree and public
BarrierRead directly in the browser, CSV downloads available; methodology is a public PDF
LicenseCC BY 4.0 — free to use with attribution
Source
From asking to doing: How the world is putting ChatGPT to workOpenAI·openai.com·2026-08-06
Site notes
All seven charts come from the original report: the two world maps are original images, and the five interactive charts were captured element-by-element by this site after rendering in a browser (the original page draws them script-side, so direct capture wasn't possible). The Chinese comparison bars and coverage diagram were created by this site. Values not labeled on the line charts were read by this site from high-resolution images and are marked "approx." in the body text; "Italy −11," "Austria −9," and "over 90% of countries" appear only on charts, never in the report text. The report text says the 35+ share was "5% higher," while the chart labels the median as "+5.1 percentage points"; this article follows the chart. Sample size, differential privacy noise, classifier methodology, age gate, population base, and CC BY 4.0 licensing all come from the OpenAI Signals methodology document (README v2.0, cdn.openai.com/signals/data-dictionary.pdf) — none of it appears in the report text. The fine-grained label composition of the six topic categories, the six age buckets, and the gender and state-level dimensions also come from that document.