Google's ATLAS Shows Where AI Use Concentrates, Not How Much

Google published an update to its AI & Economy ATLAS, a long-running research project tracking how AI is used across occupations and countries. The update has two parts: a new interactive, open-access interface for exploring what Google describes as millions of global data points, and a study produced with Google DeepMind and MIT FutureTech on how scientists use AI in their work. The post is authored by Zanna Iscenko of Google's Chief Economist's Office and Scott Strand of its Technology & Society group.

What the occupational data says

The two figures Google leads with are about concentration. In India, arts, design, and media occupations account for 19% of work-related AI usage, which Google puts at 1.6 times the global average. In the United States, computer and mathematical occupations account for 30% of work-related AI usage, double the share in the rest of the world.

The phrasing matters more than it might appear. These are shares of work-related AI usage within a country — how a country's AI use is distributed across occupations — not the proportion of people in those occupations who use AI. A creative sector can hold a large share of national AI usage because it adopted heavily, or because other sectors adopted lightly, or because of how the national occupational mix is composed. Google's own framing says India's creative industry is using AI at a higher rate than the rest of the world, which is a stronger claim than a usage share by itself supports. The distinction is the difference between "this is where AI use lands here" and "these people use AI more."

The broader occupational pattern in the data is a clean split. In OECD countries, computer and mathematical occupations plus business and financial operations lead. In non-OECD countries, the top spots go to office and administrative support; arts, design, entertainment, sports, and media; and educational instruction and library occupations. That divergence is one of the more useful findings in the release, and it argues against treating global AI adoption as a single curve that countries move along at different speeds.

Google reports that adoption generally correlates with national income level, with Brazil and the UAE standing out as higher than their GDP per capita would predict. On manual work — specifically real-time equipment diagnostics and troubleshooting — Brazil and Germany both direct 7% of work AI usage toward manual tasks, 1.4 times the global average, compared with 4% in Japan. Brazil appearing as an outlier on both measures is the kind of result worth a second look rather than a bullet point.

The methodology gap

The post does not explain where ATLAS usage data comes from. It says the project comprises millions of global data points and that the new interface lets users examine usage rates by occupation, ways people use AI at home, and adoption rates by country. It does not describe how occupations are inferred, whether the data derives from Google's own products, how self-selection and product availability are handled, or how "work-related" usage is separated from everything else.

That omission is not disqualifying — this is a blog post announcing an explorer, not a methods paper — but it constrains what a reader can responsibly do with the numbers. Cross-country comparisons of usage shares are only as good as the consistency of occupational classification and the comparability of the underlying population across markets where product availability, language coverage, and internet access all differ. Anyone planning to cite the India or United States figures in a strategy document should find the methodology before doing so.

There is also the structural point that Google is measuring AI adoption while selling AI products. That does not make the data wrong. It does mean the framing choices — which comparisons get highlighted, which get the multiplier attached — are made by a party with an interest in the direction of the story.

The science study is the more interesting half

The second component is research from Google, Google DeepMind, and MIT FutureTech that draws on ATLAS data. It combines an analysis of 2,600 specialized AI models with a survey of more than 600 scientists in the United States and United Kingdom, organized using a new taxonomy from MIT FutureTech that maps what scientists actually do.

The findings, as Google reports them: scientists use AI at a higher rate than many other occupations, with nearly half of those surveyed using some form of AI every day. Large language models and specialized models are both used widely, but for different things — LLM usage is spread broadly across scientific fields and task categories, while specialized models cluster in health and life sciences and in domain-specific prediction, generation, and simulation tasks. Scientists report saving just under seven hours a week.

Then comes the part that gives the release its credibility. Google states plainly that the time savings may not translate into discoveries. The research finds significant time spent validating AI outputs, a growing backlog of hypotheses not yet tested, and bottlenecks emerging in physical experimentation and clinical validation. The conclusion drawn is that science resembles most other occupations here: productivity gains at one stage of a pipeline do not produce output gains unless the rest of the pipeline is redesigned around them.

This is a constraint anyone who has optimized a system will recognize immediately. Speeding up a non-bottleneck stage produces queue, not throughput. If hypothesis generation gets cheap and wet-lab capacity, clinical trial slots, and expert review time do not, the observable result is a longer backlog and a validation burden that partly cancels the reported savings. The seven-hour figure is also self-reported, which tends to capture time saved on the task in front of someone rather than time added by checking the output afterward — and the same study reports that checking is substantial.

Why this framing is worth adopting

The useful contribution of this release is not any individual percentage. It is the analytical posture: measuring usage by occupation and task rather than by tool, and then asking whether measured time savings show up as measured output. Most published AI productivity claims stop at the first half.

For teams making internal decisions, the transferable method is to instrument where AI use actually concentrates in the organization, then check whether the downstream stage absorbing that output has the capacity to do so. A team that halves the time to produce analyses and does not expand review capacity has bought itself a backlog and a quality problem, not a productivity gain.

Google describes ATLAS as a long-term research project and says it will work with academic and other partners to identify new research areas. The MIT FutureTech collaboration is a reasonable start on external validity, though the taxonomy came from the partner while the usage data remains Google's.

What to watch

• Whether a documented methodology for ATLAS usage data is published, including how occupations are inferred and how cross-country comparability is established.
• Whether independent researchers can access the underlying data rather than only the visualization layer.
• Whether the science-pipeline bottleneck finding is replicated outside the United States and United Kingdom and outside a sample of 600-plus respondents.
• Whether the OECD versus non-OECD occupational split persists as AI products reach wider language and market coverage, or turns out to have tracked availability.
• Whether Brazil's dual outlier status — above-trend adoption for its income level and above-average manual-task usage — holds up in later releases.