OpenAI released country-level work-use data on October 26, 2026, showing the activities people bring to ChatGPT across nearly 100 countries. For workers, small-business owners and trainers, editing and visual-design requests offer concrete tasks around which to discuss training.
The useful question is which activities people need help with, and what they should learn to check. These figures describe requests, not completed work or demonstrated skills.
What the country comparisons show
In July–September 2026, editing written material accounted for about 12.9% of classified work-message share in India and 17.1% in the United States. Requests classified as creating visual designs or displays represented 8.7% and 7.6%, respectively. Those rounded shares come from OpenAI Signals v3.0 (2026), published under CC BY 4.0.
The percentages describe the mix of work-related messages in each country’s consumer sample—not the proportion of workers using AI. Nor do the visual-design figures establish that images were generated or usable designs delivered.
The downloadable work-activity table covers Q1 2025 to Q3 2026, with 94 countries represented overall and 74 present in every quarter. That lets readers compare the same activity across countries and follow its share over time, while keeping gaps in coverage visible.
Across those 74 countries, OpenAI reports that the average share held by each country’s ten most common activities fell from 56.0% to 51.1%. This shows a less concentrated mix of requests. It does not, by itself, establish that countries became more different from one another or that individual workers acquired new skills.
Whose work is captured?
Signals covers individual Free, Go, Plus and Pro accounts, excluding enterprise messages and Codex usage. OpenAI says this likely underrepresents business and technical uses. The population is different from the enterprise accounts discussed in RohitAI’s analysis of OpenAI’s EU agent-adoption figures.
The data dictionary describes a global sample of one million consumer messages per month from accounts declaring an age of at least 18. Deleted accounts and people who opted out of model training are excluded. Automated classifiers infer which messages concern work and assign activity categories; the published statistics include privacy-protecting noise.
Country-quarter results also need to meet a minimum sample threshold. Missing data therefore does not mean no use. These selection and classification limits make the figures unsuitable as a representative survey of a country’s workforce.
Turn task patterns into training questions
A practical reading of the data is to start with the activity, rather than a job title or a general course on AI. The following are suggested training questions, not findings that a particular course improves performance:
For editing: can staff judge whether a rewritten document preserves its meaning, names and numbers, as well as improving its tone? An exercise could compare the original with a proposed revision and require an explanation of which changes to accept.
For visual communication: can staff write a clear brief and assess whether a proposed design suits its audience and purpose? That is a different learning task from simply requesting another version.
For prioritisation: which of these activities actually recurs in the team’s work, where do people want help, and how consequential would an error be? Use those answers to choose a training example.
A national message share should not become a training-budget percentage. Repeated requests could inflate an activity’s share; a low share could reflect use of other tools or missing users. The aggregates cannot resolve those explanations. Use the country data to open a local conversation, then choose training around the work people actually do.