Most conversations about artificial intelligence still revolve around what the technology might do. It might replace workers. It might transform entire industries. It might automate everything from customer service to software development.
Google is taking a different route. Instead of asking what AI could theoretically handle, the company has examined how millions of people are already using it. The result is AI & Economy ATLAS, a large-scale research project built around nearly 15 million interactions with Google’s AI services. The first report covers users in more than 150 countries and territories, offering a rare look at what people actually ask AI to do when nobody is handing them a carefully designed experiment.
Google Built ATLAS From Nearly 15 Million AI Interactions
ATLAS stands for Activity, Task, Landscape and Adoption Study. The first version draws from 14,653,926 de-identified interactions collected between April 6 and April 19, 2026. Those interactions came from the Gemini app, Google’s AI Mode in Search and the Gemini API.
Google says it processed the information through an automated, privacy-preserving system rather than manually reading individual conversations. The dataset was aggregated and stripped of identifying information before researchers analyzed broader usage patterns.
This is not a small survey asking a few thousand participants whether they used AI last month. It is behavioral data taken from real interactions, although it still represents only Google’s own AI ecosystem. That distinction matters. The findings reveal a lot about Gemini users, but they do not automatically describe every person using ChatGPT, Claude, Copilot or smaller regional AI platforms.
The Study Covers 150 Countries, 140 Languages and Thousands of Tasks
The scale goes beyond the number of conversations. Google’s report covers more than 150 countries and territories representing around 99% of the world’s population. Researchers identified interactions in approximately 140 languages, across 800 occupational categories and roughly 4,000 different tasks. English does not dominate the dataset as much as some people might expect. It accounts for only about one-third of global AI conversations.
Users also tend to keep using their native languages when dealing with complicated requests. They are not automatically switching to English every time the task becomes technical or detailed. That is a useful reminder for AI developers. Global adoption will not be built through English-only products with a few translated menus added later. Language quality, cultural context and local knowledge are becoming central parts of the competition.
Most Gemini Conversations Have Nothing to Do With Work
One of the report’s more surprising findings is where AI use happens. More than 86% of the conversations examined were connected to personal life rather than paid employment. People used AI to research products, operate appliances, plan activities, deal with household repairs and navigate frustrating administrative processes such as taxes, licences and government fines.
These are not always dramatic use cases. Nobody is announcing a new economic revolution because someone asked Gemini how to troubleshoot a washing machine. Yet that kind of assistance may be where AI delivers its most immediate value.
Saving 20 minutes on paperwork or helping someone understand an unfamiliar repair does not necessarily appear in national productivity statistics. It still changes how people manage their time. AI adoption, at least in this dataset, looks less like one giant transformation and more like millions of small shortcuts.
AI Has Reached the Workplace, but Its Use Remains Shallow
Google found evidence of workplace AI use across professions representing more than 88% of US employment. That sounds enormous. It does not mean AI performs 88% of American work.
The study instead suggests that workers across a broad range of occupations have started using AI somewhere inside their jobs. In many professions, however, it touches only a limited portion of the tasks employees complete. This creates a strange picture: adoption is wide, but not necessarily deep.
A lawyer might use AI to summarize a document. A marketer may ask it to draft several headline ideas. An electrician could upload an image while looking for a wiring diagram. An auto technician might use it to identify an engine component. The worker has adopted AI, technically speaking. The technology has not taken over the occupation.
People Mostly Treat AI as a Collaborator
The report pushes back against the idea that most users simply hand complete assignments to an AI system and walk away. Instead, Gemini frequently acts as a collaborator. People use it for research, drafting, troubleshooting, learning and repeated refinement. Full task automation remains relatively limited, particularly when the work requires judgment or changes from one situation to another.
That pattern makes sense. A photographer might ask AI to improve an image, but each photograph still requires a different decision. A mechanic can use AI to investigate a problem, yet someone must inspect the actual vehicle. A manager may generate a first draft, then rewrite parts that sound wrong, vague or wildly confident.
The human has not disappeared. The workflow has changed. This could shift as AI agents become more capable. For now, Google’s data suggests that augmentation is far more common than complete replacement.
Skilled Trade Workers Are Using AI Too
Generative AI adoption is often framed as a white-collar trend dominated by programmers, writers and office workers.
ATLAS found more activity among skilled trade workers than researchers apparently expected. Electricians, automotive technicians and other hands-on professionals used AI to interpret images, access technical information and investigate practical problems. Workers in these fields also uploaded visual material more frequently than many office-based users.
AI may not operate the drill or replace the damaged component. It can still function as a fast reference tool sitting in someone’s pocket. That is a different form of adoption from generating emails or summarizing meetings. It may also become more important as multimodal AI improves at understanding video, equipment, physical environments and real-time camera feeds.
Wealth Still Influences Who Uses AI
AI may have spread globally, but access remains uneven. Google found that adoption generally rises alongside GDP per capita. Users in wealthier regions and higher-paid professions tend to interact with AI more often, although the report also identifies exceptions to that broader pattern. The divide is not difficult to explain.
Reliable internet access, newer devices, digital literacy, subscription costs and awareness all influence whether someone can use advanced AI tools regularly. Language support adds another layer. A platform may technically support a language without delivering the same depth or accuracy available in English. AI can spread quickly and still reproduce old inequalities.
The Dataset Has Important Limits
ATLAS offers an unusually large view of real-world AI behavior, but it is not a complete map of the AI economy. The research focuses on interactions within the Gemini app, AI Mode and Gemini API. It excludes usage inside some business-focused Google products, including Google Workspace and Gemini Enterprise, because Google does not retain the same interaction logs for those services.
That omission could affect the results. Enterprise deployments may contain more structured automation than consumer conversations. Companies often build repeatable workflows around document processing, customer support, coding, data analysis and internal operations.
The report also shows what users attempted to do. It cannot always prove that the AI response was accurate, useful or faster than completing the task another way. A conversation is evidence of use. It is not automatically evidence of success.
Google Plans to Keep Updating the Research
Google describes ATLAS as an ongoing research initiative rather than a one-time report. Future releases could show how behavior changes as AI agents gain more autonomy, multimodal systems improve and businesses connect models to internal software. They may also help researchers see whether today’s collaborative usage eventually develops into deeper automation.
That evolution will be worth watching. The first version of ATLAS does not show an economy where AI has replaced entire professions. It shows something messier. People are experimenting. They are solving small problems, drafting pieces of larger tasks and using AI in places researchers did not always expect.
Not a robot takeover. Not nothing, either. The shift is already happening. It just looks more ordinary than the headlines suggested.
