Something shifted in 2026. OpenAI’s first country-by-country usage dataset, published through their Signals platform, shows that at work people are more than twice as likely to use ChatGPT to do something — write code, edit documents, run analysis — than to ask it questions.
That ratio matters. It is the cleanest evidence yet that ChatGPT crossed from novelty into tool. Not “the future of work.” Not “a paradigm.” Just a tool people reach for when actual work needs doing.
What the data actually says
OpenAI released usage patterns for roughly a billion weekly ChatGPT users, broken down by country. Three findings stand out, and none of them are what the hype predicted.
The asking-to-doing split. During work hours, people produce. They write, edit, analyze, generate. Outside work, usage tilts toward exploration — asking questions, testing limits. The split is not subtle. It is more than 2:1 toward doing during work, which means ChatGPT is not a chatbot people goof around with. It is becoming part of how people get paid.
The geographic catching-up. Peru, Uruguay, and Costa Rica climbed the fastest in Q2 per-capita rankings. Latin America, Africa, and Oceania are closing the gap with North America and Europe — the regions that dominated early adoption. The “AI is a Silicon Valley thing” framing does not survive this dataset. Indonesia-based readers will recognize the pattern: the interesting adoption stories are no longer in San Francisco.
The age curve is bending. In almost every country, the share of messages from users over 35 grew. France and the Czech Republic saw jumps above 10 percentage points. The under-25 crowd that kicked this off is now the baseline. The growth is coming from people with mortgages.
The multimedia jump nobody talks about
Here is the number that surprised me. Multimedia — image generation, image analysis, retrieval — accounts for 7.8% of messages globally, and it is the fastest-growing category. In Brazil and Colombia, more than 1 in 10 messages involves multimedia.
That feels like an early signal. Text was the entry point because text is what large language models shipped with. Once you hand people a model that can see and draw, they use those capabilities — and they use them for work, not just play. A marketer generating ad variants does not care that image generation was bolted on later. They care that it ships the campaign.
Why the “asking to doing” shift changes strategy
If you build, market, or sell anything adjacent to AI, the asking-to-doing ratio is the number to watch. Here is what changes when your users move from asking to doing.
Retention gets stickier. Questions are disposable. You ask, you leave. Production creates artifacts — documents, code, analyses — that live somewhere. People come back to edit them. The product becomes infrastructure rather than a search bar.
Failure costs more. When ChatGPT hallucinates a fun fact, nobody loses sleep. When it ships wrong code or a miscalculated financial model into a real workflow, the stakes are different. Reliability investment stops being optional the moment people depend on output for their jobs.
Comparison shifts from “is it smart” to “does it fit.” Two years ago the benchmark was whether the model could answer a hard question. Now the benchmark is whether it slots into the work you already do without forcing you to change everything around it. Integration, not intelligence, wins the next round.
What this means if you run a team or a business
The Signals dataset is free to dig through, and it is worth your time if AI is anywhere near your roadmap. A few practical takeaways.
Stop testing AI tools with trivia questions. The signal you want is whether the tool holds up when you give it real work — your actual documents, your actual codebase, your actual analysis tasks. A model that aces a trivia benchmark but mangles your quarterly report is not ready for production.
Look at where your team already works — docs, code, spreadsheets, design files — and ask whether AI fits there, not whether your team should migrate to an AI tool. The asking-to-doing data says people adopt AI where they already produce. Meeting them there beats rebuilding their workflow around a chatbot.
Watch the age and geography data for your market. If your customers skew 35-plus, or sit outside North America and Europe, adoption is likely further along than your internal dashboards suggest. The early-adopter stereotype is falling behind the actual user base.
The boring, important takeaway
For all the coverage focused on model benchmarks and funding rounds, the most useful data point of 2026 might be this: a billion people are now using ChatGPT every week, and during work hours most of them are producing, not asking.
That is not hype. That is a habit. And habits are harder to dislodge than headlines.
Based on usage data published by OpenAI via Signals, August 2026.
