Google published the first release of its AI & Economy ATLAS study on July 23, in a post on its Google Research blog. The dataset draws on 15 million de-identified interactions spanning the Gemini app, AI Mode, and its API, three products Google says serve more than 1 billion people monthly. The sample spans 150 countries, 140 languages, 800 occupations, and 4,000 distinct tasks. Google is positioning ATLAS as the empirical foundation researchers and regulators need to understand AI’s effect on work. The company selling that AI has also chosen to be the one measuring its effects.
The findings themselves are narrower than the framing suggests. AI use touches 68% of occupations, representing 90% of total U.S. employment, according to Google’s own tally. Within a given job, though, AI accounts for only about 21% of tasks, and fewer than 10% of interactions fully automate a task rather than assist with one. Non-routine cognitive work, such as creative design and hypothesis testing, shows up in AI interactions at nearly double its share of the wider economy (65% versus 35%). Workers in manual and technical trades, including auto technicians and industrial mechanics, are twice as likely as other users to rely on multimodal features like image and video generation.
Another notable figure: more than 86% of the interactions ATLAS captured happened outside work, in tasks like researching a purchase or navigating a government form. That is a real finding. It says people are folding AI into daily life faster than into their jobs. It is also not an economic claim. ATLAS reports where conversations happen. It does not report productivity gains, wages, hours saved, or a single dollar figure tied to output or growth.
That absence matters given who is publishing the report and why. Every data point in ATLAS is proprietary telemetry from Google’s own products, processed through Google DeepMind’s internal clustering tool, OCTO. The report contains no independent estimate of AI’s contribution to GDP, employment, or income, despite branding itself as a study of “the AI economy.” A company whose valuation rests substantially on AI adoption is not a neutral instrument for measuring that adoption’s economic scale, even when its usage statistics are accurate on their own terms.
The stakes extend past marketing. Google names policymakers directly as an intended audience for ATLAS, at a moment when regulatory debates over AI in Washington and Brussels hinge on exactly the kind of claims Google is supplying: how much of the workforce touches AI, how shallow that use still is, how automation compares with assistance. When the company with the largest financial stake in a permissive regulatory environment also authors the baseline dataset that regulators will cite, the framing functions as an input to lobbying, not solely as neutral research. Google did bring in two outside academics, Diane Coyle of Cambridge and David Autor of MIT, as contributors to the report, but the underlying data, sampling method, and taxonomy remain entirely Google’s own.
Readers trying to gauge AI’s real economic footprint should treat ATLAS as a data point about Gemini’s user base, not as an account of the AI economy at large. The evidence to weigh against it is the kind Google cannot supply: peer-reviewed labor studies, government productivity data from agencies like the Bureau of Labor Statistics, and adoption surveys run by researchers with no product riding on the answer.
Published by Google Research on July 23, 2026.
