Google’s ATLAS Study Turns AI Job-Displacement Fears Into a Measurement Problem

For the past year, the loudest claims about artificial intelligence and work have converged on one alarming idea: that large language models are on the verge of automating away vast swaths of white-collar labor. Executives raise it on earnings calls, pundits raise it in op-eds, and workers raise it in survey after survey about their own job security. Google, which has more visibility into how people actually use a major AI assistant than almost anyone else, just published data that tells a far less dramatic story.

On July 23, Google Research released the first edition of the AI & Economy ATLAS (Activity, Task, Landscape, and Adoption Study), a paper built on 15 million de-identified interactions across the Gemini app, Google’s AI Mode search experience, and the Gemini API. It is among the largest publicly documented looks yet at how people use a major AI assistant in their actual jobs and actual lives, spanning more than 800 occupations, 4,000 work tasks, 150 countries, and 140 languages. The authors frame their own findings in explicitly historical terms, writing that they see “signs of a new Solow paradox,” a nod to economist Robert Solow’s 1987 quip that the computer age was visible everywhere except in the productivity statistics.

What ATLAS Actually Measured

The dataset is narrower than the “15 million” headline number implies once you look at the methodology. A joint team from Google and Google DeepMind analyzed 14,653,926 de-identified interactions collected over a two-week window between April 6 and April 19, 2026. An automated classifier first sorted each interaction into “work” or “non-work,” then a Google DeepMind clustering system called OCTO (Observation Clustering and Taxonomy Organisation) grouped related conversations and mapped them onto two established government frameworks: the Bureau of Labor Statistics’ Standard Occupational Classification paired with O*NET’s detailed task database for work activity, and the BLS American Time Use Survey lexicon for everything else. The paper positions itself alongside comparable usage studies Anthropic and OpenAI have published on their own models, adopting a similar clustering approach while extending it to a larger sample and adding household time-use analysis.

The privacy engineering shapes what the data can and cannot show. Google says original conversation text was discarded after being summarized twice, once per individual interaction and again per cluster, and any cluster representing fewer than 10 unique users was automatically excluded. Two outside academic reviewers, Diane Coyle of the University of Cambridge and David Autor of MIT, both established economists, are credited with guiding and reviewing the methodology, lending independent weight that a purely internal document would not carry on its own.

It’s just as important to be clear about what ATLAS v1.0 leaves out. By Google’s own account, the dataset excludes paid, enterprise-tier Gemini API usage through Google Cloud, along with Google Workspace, AI Overviews, Google Translate, Google Maps, Google Flow, Google Antigravity, and Gemini Notebook. In other words, the single largest blind spot in this data is precisely the category of deployment where deeper task automation would be most likely to show up first: AI wired directly into a company’s own enterprise workflows.

Broad Reach, Shallow Depth

The headline finding is a split verdict. AI use is genuinely everywhere: adoption spans over 68% of all occupations, which collectively account for just above 88% of total US employment, cutting across sectors from professional services to construction, leisure, and hospitality. That includes obviously AI-exposed roles like software developers and market researchers, but also farmers, industrial engineers, and foresters.

Reach is not depth, though. Within the median occupation that shows any AI use at all, Gemini touches only about 21% of that job’s total O*NET-defined tasks. Just 3% of occupations show AI used for more than 75% of their tasks, and Google names exactly which ones: software quality assurance analysts and testers, human resources specialists, and document management specialists. For most jobs, AI has not come close to touching the majority of what the job actually involves.

Collaboration Over Automation

Where AI does show up at work, it skews heavily toward a specific kind of task. Non-routine cognitive analytic work, tasks requiring judgment, strategic thinking, problem-solving, and creativity, makes up only 35% of all O*NET tasks economy-wide but accounts for 65% of Gemini’s work-related interactions in the dataset. Within that usage, Google’s classifier sorts most of it into four categories: partial drafting and generation, review and refinement, ideation and strategy, and information retrieval and learning. Attempts to automate a task end-to-end, with no human in the loop, make up less than 10% of AI conversations specifically within that non-routine cognitive category, not workplace AI use as a whole.

That distinction is easy to blur in a headline. The 90-percent-plus of interactions that are not full automation are still doing real work for the user: drafting a first pass, checking someone else’s output, brainstorming options, or pulling together information faster than a person could alone. That’s a genuine productivity effect. It’s just a different effect than “the AI does the job now,” which is the version of this story that tends to travel fastest.

Manual and Blue-Collar Work Isn’t Left Out

ATLAS pushes back specifically on the idea that AI is a white-collar-only phenomenon, though the data shows real limits here too. About 29% of detailed occupations show no task with meaningful AI usage at all, and Google’s own analysis finds the strongest predictor of that gap is simply how physical the occupation’s tasks are: occupations with no observed AI usage skew heavily toward manual work.

Even so, AI shows up inside plenty of hands-on trades, typically for the cognitive slice of an otherwise physical job. At Industrial Machinery Mechanics, an occupation Google notes is 44% manual by task share, thousands of recorded conversations involve workers analyzing test results and machine error messages rather than the hands-on repair work itself. Automotive service technicians and mechanics, a role that is 83% manual, generated more than ten thousand conversations in the sample about testing vehicle components, rewiring systems, and inspecting parts for wear, and that occupation’s AI conversations are more than twice as likely as the work average to involve images or video rather than text alone.

The Wage and Expertise Puzzle

ATLAS ties AI usage to pay in a way that should temper any hope that AI narrows economic inequality in the near term. A 1% increase in an occupation’s median earnings correlates with more than a 2.5% increase in AI usage intensity, a relationship that holds even after controlling for educational attainment. Weighted by actual Gemini conversation volume, the median salary across occupations in the dataset is around $83,000, roughly $20,000 above the true employment-weighted US median. In plain terms, higher-paid workers are adopting AI faster and more intensively than lower-paid ones, not the other way around.

The expertise finding is more counterintuitive. Even though the highest earners adopt AI at the highest rates overall, the specific tasks that see the most AI usage within a job tend to require lower-to-middle levels of expertise rather than a job’s hardest, highest-expertise work. Read together, the two findings are compatible: a highly paid, highly skilled worker can hand off a job’s lower-expertise sub-tasks to AI while keeping the most demanding parts of the job for themselves.

Most AI Use Isn’t Even Happening at Work

Perhaps the most underplayed number in the whole report is that over 86% of all conversational AI interactions Google measured happen outside of work entirely. AI conversations collectively touch activities that make up 98% of Americans’ waking, non-sleep hours, and Google argues this personal use generates real economic value that standard national accounts simply do not capture. As one estimate, if AI saved households just 30 minutes a week on average, the paper puts the unpaid productivity gain in the US alone at roughly $100 billion, using standard valuation methods.

Some of the heaviest non-work use clusters around exactly the kind of institutions people usually dread dealing with. Government services and civic obligations show up at almost 20 times their expected share of conversation volume, and Google found that close to half of all medical, legal, financial, and government-related AI consultations happen outside normal 9-to-5 business hours, when the actual offices are closed.

A Global Map With Notable Exceptions

Adoption tracks national wealth closely: a 1% increase in a country’s GDP per capita is associated with roughly a 0.9% increase in AI usage, and the lowest-adopting fifth of countries collectively account for only 2% of all AI conversations in the dataset, a gap Google flags as a genuine digital-divide concern. The correlation isn’t absolute, though. Several middle-income countries in Latin America and the Middle East adopt AI at rates comparable to Western Europe, and Google says the gap between countries isn’t fully explained by any single factor like internet access alone.

Language use is similarly less US and English-centric than most people would guess. English accounts for only about a third of global AI conversation volume, and work and non-work conversations show nearly identical language distributions, meaning people aren’t systematically switching out of their native language just because a task is professional. Users in non-OECD countries also generate images and video for work at roughly twice the rate of users in advanced economies.

What Google’s Own Data Doesn’t Settle

It’s worth naming the obvious tension plainly: this is a study about whether an AI company’s own product is displacing jobs, written and published by that AI company. Google has a real commercial interest in a “collaborator, not replacement” narrative landing well with regulators, the press, and enterprise customers deciding whether to expand Gemini deployments. The involvement of outside academic reviewers like Coyle and Autor, and a methodology transparent enough that outlets like Ars Technica could scrutinize it independently, tempers that concern without erasing it.

To Google’s credit, its own researchers are unusually direct about the limits of what they’re claiming. The paper states plainly that “appropriate prudence is required in interpreting these results: they constitute early observations and could evolve as the frontier of AI capabilities continues to advance, the ways in which businesses and individuals use these capabilities expand, and as the analytical methodologies economists and researchers use to measure their impact also progress.” The authors explicitly decline to rule out two of the more troubling alternative explanations still in circulation: that AI could be deepening the global digital divide, or that it could already be dampening entry-level hiring in ways a snapshot of active usage would not catch. Both remain open questions, not settled ones, by Google’s own account.

There’s also the two-week window to weigh. ATLAS v1.0 is a snapshot of April 2026 usage, not a trend line, and Google says it intends ATLAS to run as a multi-year research program rather than a one-off report. Given how quickly agentic AI tools with more autonomy have shipped over the past year, and given that this edition cannot see enterprise deployments at all, a second edition with a longer time horizon and broader product coverage could tell a meaningfully different story.

For now, though, the best large-scale usage data available, imperfect and self-interested as it may be, points toward AI as a broad but shallow presence in most jobs in mid-2026: a collaborator that drafts, reviews, and researches far more often than it replaces. Whether that holds as agentic tools mature and enterprise deployments this study couldn’t see start generating their own data is the question ATLAS was built to keep answering.