What 15 million Gemini prompts show about how workers actually use AI

Google Research examined 15 million anonymized interactions on its Gemini products and found that workers use the AI assistant for narrow slices of their jobs, not as a replacement for them. The finding, published in a paper introducing the AI & Economy ATLAS (Activity, Task, Landscape, and Adoption Study), directly contradicts the popular narrative that generative AI is about to wipe out white-collar work (Ars Technica).

Across the dataset, AI use "remains shallow and overwhelmingly collaborative in nature, with end-to-end task automation limited in scope," the researchers wrote. In practice, that means workers are pulling Gemini into discrete steps of a task: drafting a paragraph, summarizing a meeting, debugging a snippet, running a market comparison. They are not handing the assistant a full job description and walking away.

The study is the first large-scale empirical look at how generative AI is actually being used across occupations, rather than how executives or pundits claim it will be. Google built an automated classifier that maps each Gemini interaction to a Bureau of Labor Statistics Standard Occupational Classification and to O*NET's more detailed database of work activities. Human reviewers checked the probabilistic mappings and found them reliable enough to gauge workplace use at the occupational level (Google Research ATLAS PDF).

That occupational mapping produces the most useful part of the report. AI use is heavily concentrated in a few white-collar fields, and the asymmetries are sharp. Financial and market analysts, software developers, and systems administrators show up at much higher rates than the U.S. economy would predict. The same dataset shows the mirror image: salespeople, transportation workers, and food preparation and service workers barely appear at all (Google blog).

The pattern fits what the data actually measures. Software developers and systems administrators have well-defined, text-friendly tasks that match a chat interface: writing code, reading logs, drafting runbooks, summarizing tickets. Financial analysts do similar work with filings and market data. Sales calls, driving routes, and restaurant line work do not translate into prompts, and the dataset reflects that limit. The finding is as much about which jobs produce text-shaped problems as it is about which workers are curious about AI.

There is a built-in caveat the paper itself acknowledges. The data comes from Google's own products and Google's own users, mapped to Google's chosen taxonomy. The company has an obvious stake in how the AI-displacement question is framed, and the report's headline-friendly conclusion travels well inside that framing. The 15-million-prompt scale and the human-verified classifier still make this the largest public look at real AI use across occupations, but the reader should hold the result as a single-vendor snapshot rather than a population-wide measure (GCN).

For workers, hiring managers, and educators, the practical implication is that the relevant unit of analysis is the task, not the role. Most of the jobs showing heavy AI use are not shrinking; they are getting a new tool embedded into a subset of their workflows. The skills worth building sit next to the tool: prompt literacy, careful review of model output, domain knowledge the assistant does not have, and the judgment to know when to skip the model entirely. Hiring managers looking at entry-level pipelines should be planning around an AI-augmented analyst or developer, not the absence of one.

The next test is whether the same shallow, collaborative pattern holds as the underlying models get more capable and the interfaces move from chat to agentic workflows that can take actions on a user's behalf. The ATLAS dataset is a baseline. The interesting question is not whether AI use keeps growing; the data already says it is. It is whether the task share keeps widening into the parts of jobs that the current data cannot yet see.