Google AI jobs study finds Gemini use mostly assists workers

A new Google AI jobs study says workers using Gemini are mostly seeking assistance, not handing over whole jobs to the model. The finding matters because it tests broad claims about white-collar automation against a large set of real AI interactions.

Google Research released the paper last week under the name AI & Economy ATLAS, short for Activity, Task, Landscape, and Adoption Study. The team analyzed 15 million anonymized interactions from the Gemini App, Google’s AI Mode and the Gemini API, and concluded that job-related AI use is widespread but generally limited in depth.

The researchers wrote that they did not find evidence supporting claims that AI is on the verge of causing large-scale automation and displacement of white-collar work. Instead, they said current use is mostly collaborative, with full task automation still narrow.

What did the Google AI jobs study find?

Google Research classified work-related Gemini interactions using the Bureau of Labor Statistics’ Standard Occupational Classifications and O*NET’s database of job tasks. The researchers said human review found the automated classification method reliable, though some prompts required probabilistic judgments because the interactions were uncertain by nature.

Gemini use was overrepresented in several white-collar fields compared with their share of the U.S. economy, including computing, finance, arts and entertainment. Google Research identified financial and market analysts, software developers and systems administrators among heavier users for work tasks.

Other jobs showed much less Gemini activity. Sales roles, transportation workers, and food preparation and service workers were underrepresented in the data, according to the study.

To judge how far AI had spread inside particular jobs, the researchers examined whether detailed O*NET tasks appeared often enough in Gemini prompts. Across the task database, 21% of work-related tasks met the study’s threshold for “Gemini tasks,” meaning at least 25 related interactions appeared in the sample.

For 29% of occupations, no relevant work task reached that threshold. For another 30% of occupations, fewer than one-quarter of tracked tasks had significant Gemini use, suggesting that workers still handled most parts of those jobs themselves.

Only 3% of occupations had Gemini consulted regularly for at least three-quarters of the relevant tasks. Google Research listed software quality assurance analysts and testers, human resources specialists, and document management specialists among the jobs in that group.

Which tasks are workers giving to Gemini?

The study found that Gemini work use was concentrated in cognitive tasks, which made up 86% of measured interactions by volume. Interpersonal and manual tasks appeared less often than their prevalence in the workplace would suggest.

Google Research also found some use in blue-collar work. Industrial machinery mechanics used Gemini in thousands of examples involving test results and machine error messages, while auto mechanics used it in tens of thousands of conversations involving vehicle components, rewiring systems and inspecting parts for wear. Those workers were more likely than others to provide a photo rather than only text.

For cognitive work, many Gemini interactions involved drafting, idea generation, information retrieval and learning, according to the study. A smaller share related to task automation, including in work that the researchers considered routine.

The researchers said workers tended to use Gemini more for lower-expertise tasks than for the most complex parts of their jobs. Examples included rewriting material in other languages and writing or reviewing product specifications.

Google Research said the pattern may change if future AI systems improve at high-expertise work or if AI-powered robots become more capable at manual tasks. Based on current Gemini use, though, the study points to AI acting more as a tool that supports workers than as a system replacing whole occupations.

This story draws on original reporting from Ars Technica.