Google's ATLAS Study of 15 Million Gemini Chats Finds AI Assists Far More Than It Automates

Google's ATLAS Study of 15 Million Gemini Chats Finds AI Assists Far More Than It Automates

When Google set out to measure how artificial intelligence is actually being used in the economy, it did not run a survey or model a hypothetical. It read the chats. On July 23, 2026, the company published the first edition of its AI & Economy ATLAS -- short for Activity, Task, Landscape, and Adoption Study -- an analysis of 14.65 million de-identified interactions with Gemini collected over a two-week window in April. The headline finding cuts against the loudest anxieties about AI and jobs: the technology is spreading almost everywhere, but so far it is automating almost nothing.

ATLAS v1.0 draws on conversations across three surfaces -- the Gemini app, Google's AI Mode in Search, and the Gemini API -- products that together reach more than a billion people a month. Automated systems first sorted interactions into work and non-work, then mapped the work-related activity onto roughly 4,000 tasks spanning 800 occupations, more than 150 countries, and 140 languages. The clustering ran on a Google DeepMind tool called OCTO (Observation Clustering and Taxonomy Organisation), built to distill massive volumes of unstructured conversation into organized categories while stripping out personal and sensitive information. The two economists who led the report -- Zanna Iscenko of Google's Chief Economist's Office and Scott Strand of its Technology and Society team -- describe it as the opening installment of a multi-year research program, developed with input from MIT labor economist David Autor and Cambridge's Diane Coyle.

The central image is one of breadth without depth. "Adoption of AI is very, very broad in that it touches a huge range of occupations," Strand said. Gemini turned up in work tied to 68% of all U.S. occupations -- jobs that together represent roughly 90% of employment. Yet inside any given role, AI carried only about 21% of tasks on average. Strand called that pattern "very shallow." In other words, AI has a foot in most workplaces but a hand in only a slice of what each worker actually does.

Crucially, the study found little evidence of wholesale task automation. The overwhelming majority of work interactions were collaborative -- ideation, strategy, information retrieval, drafting, troubleshooting, and learning. Tasks Google classifies as "non-routine cognitive," such as creative design and hypothesis testing, showed up in AI work conversations far more often than their share of the broader economy (65% versus 35%). But fewer than 10% of those interactions appeared aimed at fully automating the task. People, in short, are using Gemini as a collaborator, not a replacement.

The data also punctured a few assumptions. Blue-collar and technical workers used AI more than researchers expected: auto technicians reading test results, electricians chasing wiring diagrams, mechanics inspecting machinery. "Blue-collar workers tend to be using a lot of what we call multimodal AI, which is AI with images and video," Strand said, noting they were twice as likely to lean on those modes. And most usage was not about work at all: 86% of Gemini interactions happened outside the job, from cooking and appliance help to navigating government services like taxes and licensing. Wealth still shapes the map -- a 1% rise in an occupation's median earnings was associated with a 2.68% increase in AI use -- but adoption already spans countries covering 99% of the world's population, with English accounting for only about a third of conversations.

The findings land in an increasingly crowded field. Anthropic's Economic Index, first published in early 2025, similarly sorted Claude conversations into "augmentation" and "automation" and found augmentation in the majority -- but it also reported a meaningfully higher automation share and estimated that roughly 36% of occupations were using AI for at least a quarter of their tasks. OpenAI's own labor analysis pointed in a comparable direction. Google's numbers are directionally aligned with both, while landing on the more conservative end of the automation question. That the three largest model makers are now publishing convergent, if not identical, pictures of a mostly augmentative present is itself notable.

The caveats are real, and Google names several. ATLAS deliberately excludes Google's business tools -- Gemini Enterprise and Workspace -- because the company does not keep the necessary logs, and enterprise settings are precisely where deeper automation might be taking hold. "We obviously would love to use that data if we could," Strand said. "It just wasn't even an option for us." There is also the obvious tension that the measurement is coming from the company selling the product being measured, and that a two-week snapshot captures how people have used AI, not where the trend is headed. The involvement of independent academics like Autor and Coyle is Google's hedge against that skepticism.

What to watch is whether ATLAS becomes a genuine time series. A single reading showing shallow, assistive use is reassuring; a rising automation share across successive editions would tell a very different story. Google says ATLAS is the first in a planned series but has not committed to a publishing cadence. For now, the most comprehensive public look at real-world AI use suggests the revolution, at work, still mostly has a human in the loop.