Editorial Team
7 min read
Google’s first ATLAS report finds Gemini use across occupations covering 88.4% of U.S. employment, but mostly in a limited share of each job’s tasks. The data offer a detailed view of how Google’s AI is used; they do not show whether it is raising output, eliminating work or changing hiring.
Google has released a large new map of how people use Gemini. It is not, despite the inevitable labor-market reading, a measure of whether artificial intelligence is replacing workers.
The company’s AI & Economy ATLAS v1.0 analyzes 14,653,926 de-identified interactions from April 6 through April 19, 2026. It samples Google’s Gemini app, AI Mode search experience and Gemini API, then uses automated systems to classify activities against U.S. job and time-use taxonomies. Gemini is Google’s generative-AI product family; the app and AI Mode are consumer conversational products, while the API lets developers build Gemini into software.
That scope makes ATLAS an unusually detailed product-usage study. It also defines its limit. The report records what users asked Google’s systems to do, not whether a task was completed, an answer was correct, time was saved or an employer changed staffing.
Google’s company-reported task-intent classification by task category. Source: Google AI ATLAS report.
The study observed meaningful Gemini use in 68% of detailed occupations. Those occupations account for 88.4% of U.S. civilian employment, a coverage calculation based on U.S. employment data—not a claim that 88.4% of U.S. workers use Gemini.
Depth was far lower. Among occupations with at least one task meeting the report’s usage threshold, the median task saturation was 21%; only 3% had observed use in more than three quarters of their listed tasks. The threshold itself matters: an occupation counted as covered needed at least 50 users globally on a task, while task saturation required at least 25 users globally for a listed task.
This is evidence against treating occupational exposure as job replacement. It is not evidence that automation is absent. In the report’s task-intent classification, less than 10% of interactions involving non-routine cognitive work—such as analytical problem-solving or creative design—were requests to automate a task end to end. The company classified more than a quarter of routine cognitive interactions that way.
The distinction is consequential. A headline average can obscure changes in a narrower set of repeatable workflows. And an automation request is still not an observation that a worker’s task, let alone a job, disappeared.
Google’s researchers route conversations through several model-based steps: personal information is redacted; the data are classified as work or non-work; conversations and then clusters are summarized; and the clusters are assigned occupations, tasks and intent. Clusters representing fewer than 10 users are discarded, and original conversations are not retained in the final dataset.
The privacy design limits what the researchers can inspect, but it also creates a chain of inference between a prompt and a labor-market claim. In a synthetic validation test, the work/non-work classifier reached 93.7% accuracy. Exact assignment to one of 18,797 O*NET tasks reached 22.58%, although the broader five-category task grouping used for much of the report reached 70.44% accuracy. Google says its more aggregated findings are therefore more robust than granular occupational labels.
The data are also not a representative census of work. The report excludes Google Workspace, AI Overviews, Translate, Maps and other Google products. More importantly for workplace automation, paid Gemini API requests and enterprise use through Google Cloud lack the request content required for the detailed pipeline; Gemini Enterprise and Workspace logs are not included. The company says it plans to expand coverage in later editions.
That omission cuts both ways. Consumer and free-product activity cannot settle how businesses are redesigning work, and the excluded systems could contain more automation. But it would be just as unwarranted to assume that they do. ATLAS can describe the included Google surfaces, not the whole AI market or the labor market.
The report’s most useful contribution may be to separate task assistance from physical execution and from labor-market outcomes.
For manual and technical work, Gemini often appears as an information tool rather than a substitute for the physical job. Automotive service technicians and mechanics used it for vehicle tests, rewiring and inspection; their multimodal conversations were more than twice as common as the overall work baseline. The report also finds manual and interpersonal tasks under-represented in Gemini activity compared with the O*NET task universe.
The distribution is uneven across income and geography. For U.S. occupations in the study, a 1% increase in median earnings was associated with 2.68% higher Gemini-use intensity in a simple regression. When the report included educational attainment as well, the earnings association was 1.86%. These are occupational correlations—usage intensity is classified conversations divided by 2024 employment in that occupation—not proof that higher pay causes AI use or that AI raises pay.
Globally, Google adjusted conversational-use estimates using a Gemini web-referral proxy and found a 0.9% association between GDP per capita and conversations per capita for each 1% increase in GDP per capita. Its lowest-use country quintile represented roughly 17% of the population in the analysis but 2% of conversations; the highest-use quintile represented 11% of the population and 30% of conversations. The report cautions that platform reach, internet access, local alternatives, language and occupational mix can all influence those figures.
For the Gemini app and AI Mode, excluding the API, 86.5% of sampled conversations were classified as non-work. Leisure and socializing accounted for 26.4% of total conversational use and education for 20.7%.
Google also found government services and civic obligations almost 20 times as prevalent in U.S. AI conversations as their share of non-work time in the American Time Use Survey. That can describe where people seek help; it does not establish the quality of the help. The report does not know whether users followed an answer or resolved the underlying legal, medical, financial or administrative problem.
Its household-value estimate should be read in the same light. Using an average 17.8 hours a week of defined productive household work, a $12.02 hourly replacement wage and assumed AI time savings of 0.5% to 5%, the report estimates $14.9 billion to $149 billion a year in possible U.S. value. The middle 2% scenario is just under $60 billion. Those are scenarios, not measured savings: the authors explicitly say they have no causal evidence for how much time AI saves at home.
ATLAS was written by Google and Google DeepMind researchers, including Zanna Iscenko, AI & Economy lead in Google’s Chief Economist’s Office, and Scott Strand, head of strategy and special projects in Google’s Technology & Society organization. Google is both the researcher and the operator of the products whose usage it is measuring, giving it unique data access and a direct stake in how the findings are interpreted.
The company says University of Cambridge economist Diane Coyle and MIT labor economist David Autor provided contributions, guidance and review. Coyle is the Bennett Professor of Public Policy and research director of Cambridge’s Bennett School; Autor is an MIT professor and co-director of the National Bureau of Economic Research’s Labor Studies program, and is a three-year visiting fellow in Google’s Technology & Society program, according to the program’s description. Their involvement adds relevant economic expertise, but it does not make the underlying dataset independent: Google collected, processed and analyzed it.
The report is part of Google’s AI & Economy Research Program, which says it works with academic and other researchers on AI’s effects on work, productivity and economic transformation. That context helps explain both the project’s ambition and why independent outcome data are needed alongside it.
The next edition’s most important test is not whether it finds more prompts. It is whether it can connect usage to independently observable results: completed work, error rates, hours, output, pay, hiring and displacement.
Broader enterprise coverage would help reveal whether organizations—not just individual users—are removing, redesigning or adding tasks. Transparent tests of the classifiers against real, consented work records would help establish whether the labels retain their meaning outside synthetic validation. Until then, ATLAS is strong evidence that Gemini has diffused broadly and is used selectively. It is not evidence that AI’s labor-market consequences have been measured, much less resolved.
AMD has introduced ROCm.AI to guide setup, deployment and optimization on its hardware, while HIP and HIPIFY offer a separate source-level route from CUDA for embedded software. The unresolved question is whether those tools lower enough engineering risk to matter for large planned AI deployments.
KTransformers v0.6.4 adds LoRA and hybrid fine-tuning to its CPU-GPU workflow for sparse mixture-of-experts models, expands an Ampere inference path and moves a scheduler endpoint to loopback after a report of unauthenticated pickle deserialization. Its published throughput figures are configuration-specific release claims, not a general comparison of fine-tuning methods.
Microsoft has put MAI-Image-2.5-Pro and MAI-Voice-2-Flash into public preview in Foundry. The launch extends the company’s in-house AI options, but its evidence for lower cost and broader deployment is chiefly tied to MAI-Image-2.5 and company-reported workloads—not an independently comparable test of the new image model.
Microsoft has made its MAI-Code-1-Flash coding model available in paid GitHub Copilot plans and says a related model is live in Excel. Its evidence points to a targeted effort to reduce serving costs, while leaving the scale, task coverage and economics of the broader rollout unresolved.
AMD has detailed the 256-core EPYC 9996, its density-oriented Venice server CPU, alongside a broader 9006 platform with 16 memory channels and PCIe 6.0. But the Helios AI rack uses a separate high-frequency Venice design, and AMD’s largest rack-performance figures remain preliminary company projections rather than independent results.
NVIDIA says it will commit $4 million over three years to let CNCF projects test on real GPUs. With Kubernetes’ core Dynamic Resource Allocation framework already generally available, the consequential question is whether shared validation can make diverse GPU implementations dependable without settling operators’ choices over isolation and control.
Microsoft has made MAI-Image-2.5 the end-to-end default for Bing Image Creator and put it into PowerPoint image-to-image features. The company says the PowerPoint deployment can reduce GPU costs by up to 84% versus GPT-Image-2, while its much larger OpenAI partnership remains in place.
Centrica says its roughly 1,300 proposed role reductions reflect lower customer contact and a wider transformation programme. The record shows AI is one tool in that programme, while the headline total reaches beyond the 500 contact-based roles it has specifically identified.
Meta is using a new mass-market film to argue that AI will extend its mission of connection. The company has real distribution, product and infrastructure plans behind the pitch, but the meaningful test is whether it makes access, controls and outcomes legible as those plans reach users.
Google is rolling out an optional selfie-video check for eligible Google Account holders who lose access to their usual devices. The feature adds a biometric reference to recovery, but it excludes several account types and leaves users dependent on Google’s unreported real-world performance and recovery decisions.
Nokia booked €2.8 billion in AI and cloud order intake in the second quarter, as optical and IP-network sales grew. The orders strengthen its data-centre strategy, but they are not revenue yet, and restructuring, supply constraints and negative cash flow leave execution as the central question.
Super Micro Computer says it received more than $60 billion in fiscal fourth-quarter orders and now expects 15% to 17% gross margins, despite revenue tracking near the low end of its outlook. The preliminary figures point to a much larger AI-server delivery pipeline, but not yet to booked revenue or a durable improvement in profitability.
Nebius says it has brought up its first full NVIDIA Vera Rubin NVL72 rack in Finland. The deployment is an early integration milestone backed by a $2 billion NVIDIA private placement, but the company still has software and production testing to complete before customers can use it.
According to people familiar with the effort, Amazon founder and executive chairman Jeff Bezos has pressed Prime Video to make AI and personalization central to a proposed redesign known internally as Lighthouse. The early test could improve discovery, but Amazon has not said how it would reconcile personalized rankings with the paid placement and subscriptions that make Prime Video a complex storefront.
A report says some Chinese cities are gradually issuing robotaxi permits again after a review prompted by Baidu’s Wuhan outage. Shenzhen’s revised rules show that road testing, demonstrations and unmanned trials remain separately governed—and do not by themselves establish broad commercial service.
Databricks says it will run core operations and analytics on Azure Databricks, deepen Microsoft product integrations and expand use of Azure Cobalt chips. The partnership runs into the 2030s, but the companies have not disclosed its value, workload scale or customer results.
OpenAI is beginning a U.S. rollout of Health in ChatGPT for logged-in adults on web and iOS. The feature can bring optional Apple Health and supported medical-record context into ordinary chats, but connected data may be incomplete and requires user permission before use.
Yelp will supply ChatGPT with reviews, ratings, photos and business details, with branded links back to Yelp and a planned quote-request feature. The non-exclusive deal extends an existing licensing strategy, but neither company has disclosed its financial terms or how ChatGPT referrals will be measured and sold.
U.S. stocks fell as attacks on Saudi tankers added a second threat to Middle East oil flows and earnings from Alphabet and Tesla highlighted the cash demands of the AI buildout. The selloff was concentrated in megacap technology rather than a uniform retreat from equities.
Blackstone’s second-quarter distributable earnings rose 26% as realizations and fee-related earnings increased. Its data-center and other AI-linked investments are becoming a larger part of the firm’s portfolio narrative, but the results do not separately show how much of the gain came from AI while retail private-credit fundraising slowed.