What Work Does Generative AI Do?

Since August 2024, our Real-Time Population Survey (RPS), a nationally representative survey of U.S. adults ages 18 to 64, has tracked how many Americans use generative artificial intelligence (AI). The adoption of generative AI — hereafter simply “AI” — keeps climbing: Between August 2024 and May 2026, the share of adults using AI rose from 45% to 62%, and the share of workers using it for their jobs rose from 33% to 45%.

But “how many” workers use AI is only one factor in determining the technology’s effect on the labor market. The impact of a new technology on employment and wages for different groups also depends on what types of work it can do. Work that a technology can take over tends to lose value in the labor market; work that it cannot do — but that still must get done — tends to gain value. This means that understanding how AI is changing the labor market requires first knowing what work it actually does. Our new working paper, summarized in this blog post, presents this information using the first nationally representative measures of AI adoption at the level of detailed work tasks.

How Do We Measure AI Use at the Task Level?

Across the economy, workers perform a huge variety of job tasks. One common framework is O*NET, a Labor Department-sponsored database, which contains a list of roughly 2,000 detailed work activities — everything from “prepare research reports” to “drive trucks or other vehicles to or at work sites.” It is simply infeasible to ask RPS respondents to identify from such a long list which job tasks they perform and, moreover, for which of them they use AI.

Our solution to this challenge is to take advantage of another useful feature of O*NET: It rates how important each of these 2,000 activities is for a given occupation. In the RPS, we first elicit the respondents’ occupations. Next, we show the respondents the 10 most important tasks in their respective occupations, according to O*NET, and ask which ones they perform. Finally, we ask those who use AI at work to report the tasks AI tools regularly help them complete.

These data allow us to compute adoption indexes that answer two different questions. First, among workers in a given occupation, what share use AI? Second, among the workers who perform a given work task, what share use AI for that task? Our adoption indexes, based on nearly 14,000 workers surveyed across four quarterly waves between August 2025 and May 2026, reveal important insights about the state of AI adoption in the U.S. The data are publicly available for download.

AI Adoption Is Widespread but Shallow

The figures below plot the distribution of AI adoption rates across detailed occupations and detailed work tasks.

These figures carry two messages:

First, AI adoption is widespread. In more than 80% of occupations, at least 1 in 5 workers uses AI on the job, and more than 40% of tasks have adoption rates above 20%. AI is not confined to a handful of tech jobs — it is assisting a broad share of labor market activity.
Second, AI adoption runs shallow almost everywhere. Only 40% of occupations have adoption rates above 50%, and just 16% exceed 70% adoption. Tasks are starker still: Fewer than 3% of tasks have adoption rates above 50%, and none exceed 70% adoption. One lesson is that AI use in the highest-adoption occupations is not driven by a single task for which everyone employs the technology; instead, in those occupations, nearly everyone uses AI, but workers use it for different parts of their jobs.

Where Is AI Adoption Highest and Lowest?

The ranking of occupations by AI use is not all too surprising. Adoption is highest in computing and professional occupations: computer and information research scientists (87.3%), information security analysts (85.4%), and network and computer systems administrators (82.4%). Computer programmers, public relations specialists, personal financial advisors and chief executives are all near or above 80%. The most-assisted tasks are cognitive and information-intensive: reading documents to gather technical information (61.3%), preparing research reports (60.7%) and analyzing data to identify trends (57.5%).

Adoption is lowest in occupations where work is hands-on or face-to-face: animal caretakers (5.3%), receptionists and information clerks (7.6%), and licensed practical and licensed vocational nurses (10.4%). The task data show why. Among the tasks with no reported AI use are driving trucks, presenting menus to customers, assisting practitioners with medical procedures, collecting biological specimens from patients and preparing treatment areas. These results partly reflect conventional wisdom: Currently, AI works on screens and cannot drive a truck or draw blood.

What Predicts AI Adoption?

A popular research approach ranks occupations and tasks by their “exposure” to AI. Do our adoption data simply reinforce existing exposure predictions?

The answer is partly yes and partly no. Exposure scores do reasonably well at predicting which occupations and tasks exhibit higher adoption rates.
For example, some exposure predictions explain roughly half of the variation in AI adoption across occupations and tasks.

However, we document several cases in which predicted exposure and actual adoption sharply differ. Occupations built around sensitive records adopt far less than predicted; for example, medical secretaries and administrative assistants use AI at a 16.8% rate versus a predicted 61%. The same is true of tasks like prescribing medical treatments or examining documents for accuracy and compliance, for which privacy rules and the cost of errors loom large.

In the other direction, computer and office machine repairers (75.7%), special education teachers (70.9%), and even laundry and dry-cleaning workers (49.0%) adopt AI at roughly double the rate predicted by common exposure measures. These jobs require manual or face-to-face work, but they also require information-gathering, planning and communication that AI can assist. The same holds for job tasks such as researching laws and legal precedents or developing business and marketing plans. When workers have latitude over how their work gets done, they often seem to find uses that the scores did not anticipate.

Finally, while exposure scores do reasonably well in ranking adoption rates across occupations or tasks, they are poor predictors of whether a given individual adopts AI. The figures above show why: Within most occupations and tasks, some workers adopt AI and most do not, so knowing what a worker does — which is all an exposure score reflects — is not all that informative about whether that particular worker adopts AI.

What does tell us a great deal is who the worker is. Specifically, we find that some workers systematically use AI for more tasks than others, even when we compare workers performing similar tasks. Demographic characteristics such as age, education and sex explain little of these differences. Instead, our evidence suggests that an important determinant of AI adoption is learning from experience. For example, workers who have used generative AI for at least six months adopt it for more of their work tasks, and workers whose job tasks make them likely adopters are also more likely to use AI outside of work. This evidence is consistent with a story of costly learning or experimentation: Once workers have invested in learning how to use AI in one domain, they are more likely to use it elsewhere. This mechanism suggests AI’s footprint will continue to deepen as more workers become familiar with the technology and use it in more areas of their work.

Looking Ahead

Our occupation- and task-level adoption indexes provide a new input for studying the effects of AI on productivity, wages and employment. Our findings also carry a broader lesson: Understanding why some workers adopt AI while others do not may be as important as understanding what the technology can do. As adoption continues to deepen, the indexes will let us track not just how many workers use AI, but what work it does.

Notes

See the September 2024 On the Economy blog post “The Rapid Adoption of Generative AI,” the February 2025 blog post “The Impact of Generative AI on Work Productivity,” and the November 2025 blog post “The State of Generative AI Adoption in 2025.” Current adoption rates are available from our AI adoption tracker and on FRED.
See, for example, the 2003 study “The Skill Content of Recent Technological Change: An Empirical Exploration” by David H. Autor, Frank Levy and Richard J. Murnane.
See our 2026 working paper, “What Work Does Generative AI Do?” Chat logs from generative AI platforms offer an alternative window into what work AI does, but they come with their own conceptual and measurement challenges. For example, chats are typically classified into tasks without knowing the user’s occupation. We discuss these challenges in the working paper but not in this blog post.
The occupation- and task-level AI adoption indexes can be downloaded through the RPS. They will be updated with future survey waves.
An occupation’s adoption rate is the share of workers in that occupation who report using AI for their jobs. Occupations are defined using three-digit codes from the Bureau of Labor Statistics’ Standard Occupational Classification system. A task’s adoption rate is the share of workers who perform the task and report using AI for it. Tasks are those listed among O*NET’s detailed work activities.