A new Google AI jobs study gives the hype machine a useful speed bump: workers are using Gemini across many occupations, but Google Research says the data does not show broad replacement of white-collar labor. In a paper released last week, the company’s researchers found that workplace use of Gemini is mostly assistive, with full task automation still limited.
The paper introduces Google Research’s AI & Economy ATLAS, short for Activity, Task, Landscape, and Adoption Study. The project examined 15 million anonymized AI interactions from the Gemini app, Google’s AI Mode, and the Gemini API.
That matters because the study looks at what people actually asked Google’s models to do, rather than what executives, investors, or demo videos claim AI will do someday. It is still Google studying Google usage, with all the limits that implies. But the dataset is large enough to make some of the louder claims about imminent office-worker obsolescence look undercooked.
What did Google measure?
Google Research classified work-related Gemini interactions using the Bureau of Labor Statistics’ Standard Occupational Classifications and O*NET, the U.S. database that breaks jobs into specific tasks. The researchers used an automated classifier, and they said human reviewers checked the method and found it reliable, even though classifying prompts involved uncertainty.
The results were uneven across the labor market. Gemini use was more common than expected in some office-heavy fields, including computing, finance, arts, and entertainment. Financial and market analysts, software developers, and systems administrators were among the heavier users for job tasks.
Other occupations barely showed up relative to their size in the economy. Google Research said sales workers, transportation workers, and food preparation and service workers were underrepresented in the usage data.
Does Google’s AI jobs study show workers are being replaced?
No, at least not at the scale suggested by some automation claims. Google Research found that only 21 percent of all work-related tasks in the O*NET database met its threshold for meaningful Gemini usage. The researchers counted a task as a “Gemini task” when the dataset contained at least 25 related interactions.
For 29 percent of occupations, no tracked task reached that threshold. Another 30 percent of occupations had meaningful Gemini use in less than one-quarter of their listed tasks. That points to limited penetration inside most jobs, rather than software quietly swallowing entire roles.
The most exposed slice was small. Google Research said only 3 percent of occupations had regular Gemini consultation for at least three-quarters of relevant tasks. Software quality assurance analysts and testers, human resources specialists, and document management specialists were examples in that group.
What are workers asking Gemini to do?
Most Gemini workplace usage involved cognitive work. Google Research said 86 percent of measured interactions by volume fell into that category, while interpersonal and manual work appeared less often than those tasks do in the broader workplace.
Manual jobs were not absent. The study found industrial machinery mechanics using Gemini to interpret test results and machine error messages. It also found many auto mechanic conversations involving vehicle testing, rewiring, and inspection tasks. Those users were more likely than others to include photos rather than rely only on text.
For knowledge work, the common requests clustered around drafting, generating ideas, finding information, and learning. A smaller share involved automating job tasks, including tasks Google categorized as routine.
The researchers also found that workers tended to use Gemini for lower-expertise cognitive tasks, rather than the hardest parts of their jobs. Examples included translation-style rewriting and work around product specifications.
Google Research said the current pattern looks more like augmentation than mass displacement: employees use AI to handle routine cognitive chores and to collaborate on some non-routine work. The paper also cautions that this could change if future models improve on high-expertise tasks or if AI-powered robots become better at manual work. For now, the data says the boring part out loud: adoption is real, but the robot pink-slip factory is not here yet.
This story draws on original reporting from Ars Technica.
