The Rapid Adoption of Generative AI

Source: NBER Working Paper 32966 (Alexander Bick, Adam Blandin, David J. Deming) · September 2024, revised February 2025

Generative artificial intelligence (AI) is a potentially important new technology, but its impact on the economy depends on the speed and intensity of adoption. This paper reports results from a series of nationally representative U.S. surveys of generative AI use at work and at home. As of late 2024, nearly 40 percent of the U.S. population age 18-64 uses generative AI. 23 percent of employed respondents had used generative AI for work at least once in the previous week, and 9 percent used it every work day. Relative to each technology’s first mass-market product launch, work adoption of generative AI has been as fast as the personal computer (PC), and overall adoption has been faster than either PCs or the internet. Generative AI and PCs have very similar early adoption patterns by education, occupation, and other characteristics. Between 1 and 5 percent of all work hours are currently assisted by generative AI, and respondents report time savings equivalent to 1.4 percent of total work hours. This suggests that substantial productivity gains from generative AI are possible.

Introduction. Generative artificial intelligence (genAI) has rapidly emerged as a potentially important workplace technology. The large language model ChatGPT debuted in November 2022, and by December 2024 it reported over 300 million global weekly users. Several recent studies find that genAI improves worker productivity (Brynjolfsson, Li, and Raymond, 2023; Cui et al., 2024; Dell’Acqua et al., 2023; Noy and Zhang, 2023; Peng, Kalliamvakou, Cihon, and Demirer, 2023; Toner-Rodgers, 2024) and some economists project that genAI will produce large aggregate economic effects (Chui et al., 2023; Filippucci, Gal, and Schief, 2024). Yet other studies predict only modest impacts of genAI on employment and productivity (Acemoglu, 2024). Uncertainty over the impact of genAI on the economy in part reflects a lack of systematic evidence on the frequency and intensity of genAI adoption, and most studies pre-date the release of widely accessible genAI tools (Acemoglu, Autor, Hazell, and Restrepo, 2022; Albanesi et al., 2024; Gathmann, Grimm, and Winkler, 2024).1 This paper studies the adoption of genAI using data from the first nationally representative U.S. survey of genAI usage at work and at home. Our data come from the Real-Time Population Survey (RPS), a national online labor market survey of working age adults aged 18-64 that has run since 2020 (Bick and Blandin, 2023). We report the combined results of two surveys fielded in August and November 2024, which collectively included more than 10,000 respondents.2 The RPS asks the same core questions and follows the same timing and structure of the Current Population Survey (CPS). This parallel structure allows us to benchmark our survey estimates against the CPS and to construct weights that ensure a nationally representative sample. Benchmarking to the CPS also allows us to compare the adoption of genAI to the personal computer (PC) and the internet. Beginning in 1984 the CPS fielded the Computer and Internet Use (CIU) supplement, which fueled an influential literature studying the impact of computerization on the labor market (Autor, Katz, and Krueger, 1998; Card and DiNardo, 2002; Krueger, 1993). We closely follow the CIU question wording and structure, which allows us to compare adoption of different technologies relative to their first mass market product release. But we also introduce novel questions to measure how intensively individuals use the technology and how much time it saves them.

We report six main results. First, a substantial share of respondents already use genAI at work and at home. In August and November of 2024, 26 percent of employed respondents reported using genAI for work: 9 percent used it every work day, 14 percent on some but not all work days in the previous week, and 3 percent did not use it in the last week. One third of respondents report using genAI outside of work, with 27 percent using it at least once in the previous week. Overall, 39 percent of respondents report using genAI either for work or outside of work. The most commonly used products are ChatGPT (28 percent of all respondents), Gemini (17 percent) and embedded products such as Microsoft Copilot (14 percent).

The widespread adoption of computers and related information technologies beginning in the 1980s contributed to rising aggregate productivity and growing income inequality (Autor, Levy, and Murnane, 2003). Comparing early adoption patterns of genAI to computers can help us understand whether genAI will have a similar impact to these earlier technologies (Autor, 2024). Our second main finding is that the adoption of genAI has so far been at least as fast as adoption of computers and the internet. In 1984, 25 percent of workers reported using computers for their job, compared with 27 percent of workers who report using genAI for their job in 2024. (The year 1984 was three years after the release of the IBM PC, the first massmarket computer; from this perspective, workplace adoption of genAI in 2024, two years after the release of ChatGPT, can be viewed as being on a similar trajectory.)

In contrast, genAI has been adopted at a faster rate than computers outside of work. While we cannot distinguish between work and non-work internet use, overall genAI adoption (including work and non-work usage) has been faster than both the personal computer and the internet. This could be due to differences in cost and user-friendliness. Computer adoption required the purchase of expensive hardware, and the internet required the purchase of a modem and an internet service provider contract, whereas ChatGPT is available for free or with a lowcost subscription, and using it does not require any technical expertise.

Method. Our main data source is the Real-Time Population Survey (RPS), a national labor market survey of U.S. adults aged 18-64 (for a detailed discussion, see Bick and Blandin 2023). The RPS is fielded online by Qualtrics, a large commercial survey provider, and has collected multiple survey waves each year starting in 2020.

The RPS is designed to mirror the Current Population Survey (CPS) along key dimensions. The RPS survey matches questions on demographics and labor market outcomes in the basic CPS and CPS Outgoing Rotation Group, using the same word-for-word phrasing when practical and replicating the intricate sequence of questions necessary to elicit labor market outcomes in a manner consistent with the CPS (US Census Bureau, 2015). Replicating key portions of an existing high-quality survey ensures that survey concepts are comparable, which allows researchers to validate RPS outcomes against a widely used benchmark with a larger sample size and, where necessary, to construct sample weights. Bick and Blandin (2023) show that the RPS produced very similar statistics for employment, hours worked, earnings, industry composition, and employee tenure during the pandemic.4 The RPS also collects information not contained in the CPS, including trends in employee reallocation across firms, work from home, and interstate migration as well as the relationship between inflation and job search (Bick and Blandin, 2023; Bick, Blandin, and Mertens, 2023; Bick, Blandin, Mertens, and Rubinton, 2024; Pilossoph and Ryngaert, 2023). Beginning in June and continuing through November 2024, the RPS introduced a module designed to measure genAI use both at work and at home.

2.2 Measurement of Generative AI Use A key goal when designing our survey was to provide a comparison to the historical adoption of other technologies. To facilitate this comparison, we use existing technology adoption questions from the CPS as a template for our own questions regarding genAI. In October 1984, the CPS launched a supplemental module to study the adoption of a new technology: the computer. (It was eventually named the Computer and Internet Use supplement, CIU). The leading question about computer adoption at work for employed respondents was:

Because genAI is a relatively new technology, we believe it is important to provide both a definition of the concept and some specific examples to our survey respondents. We therefore start the genAI survey module with a definition of genAI:

Generative AI is a type of artificial intelligence that creates text, images, audio, or video in response to prompts. Some examples of Generative AI include ChatGPT, Gemini, and Midjourney.

We included examples of popular genAI products because we thought some respondents may be more familiar with those product names than with the broader concept of genAI. After defining genAI, the module asks respondents whether they had heard of the concept prior to the survey. Overall, 74.5% report having heard of genAI; the remaining 25.5% who answered “No” skip the remainder of the module. Respondents who answer “Yes” continue with the survey, which then later asks them to select from a list of products or to write in others that they have used.

For employed respondents, the next question asks about genAI use at work:

Do you use Generative AI for your job? (No/Yes) This question is designed to mirror the analogous computer use question from the CIU, discussed above. We opted for the version asked from 2001 onwards that omitted the word directly.

A second question—though not immediately followed by this question—asks about genAI usage outside work.

Do you use Generative AI [outside your job]? (No/Yes) Non-employed respondents are not shown the term in brackets. While the CIU asks about computer use “at home”, we do not use this language to account for respondents who access genAI on mobile devices outside of their home.

All genAI users are asked about which product they used and can select from a list of products (e.g. ChatGPT, Google Gemini) or to write in others that they have used.

Discussion. Higher wages by more intensive users could reflect several factors. First, there is substantial variation in job tasks even within occupation and industry, and tasks that benefit more from genAI assistance may be associated with higher wages (Deming and Kahn, 2018). Second, productive workers may adopt new technologies more rapidly (Nelson and Phelps, 1966). Third, genAI may directly increase workers’ productivity and wages. Our data do not allow us to distinguish between these explanations.

Figure 7 compares early adoption patterns for the two technologies by occupation and industry. Figure 7a plots genAI adoption against PC adoption by occupation group.13 genAI adoption at work is highest in computer and mathematical, management, and business and finance jobs (43, 42, and 41 percent respectively). However, genAI use is broadly distributed, with every occupation group between 15 and 50 percent. In contrast, PC use is more concentrated, with three occupation groups above 50 percent and two occupation groups below 10 Our results offer partial validation of task-based exposure predictions, supporting the utility of these measures for researchers. At the same time, managers have relatively high adoption rates relative to predicted exposure, while legal and social services and office and administration occupations have relatively low adoption rates. These disparities could reflect some mismeasurement in which tasks can benefit from genAI. Alternatively, they could reflect variation in adoption costs or regulatory barriers.

The close correspondence between predicted exposure based on occupational tasks and actual adoption suggests that the task content of a worker’s job may be an important determinant Using November 2024 RPS data on hourly wages wi, weekly hours worked li, and weekly genAI time savings si, we use (7) to estimate that genAI currently increases aggregate labor productivity by 1.9%.19 If we assume an AI-exposure-adjusted labor share of 0.57, following Acemoglu (2024), the implied potential productivity gain is 1.1%.

However, labor productivity gains will only translate into aggregate productivity gains if firms adjust their expectations. If workers are now able to complete the same tasks in less time without their employer’s knowledge, they may take their time savings as on-the-job leisure, which would increase welfare but not productivity. Bonney et al. (2024) find that only 5.4 percent of firms had formally adopted genAI as of February 2024. While firm adoption may increase over time, it still lags far behind, suggesting that worker adoption is still mostly informal. Thus we emphasize that these potential productivity gains from genAI may not show up in productivity statistics, at least for now.20 Our estimate of the potential aggregate productivity gain from genAI of 1.1% is slightly higher than the estimate of 0.7% found by Acemoglu (2024). Our estimate is based on reported genAI use and time savings, while Acemoglu (2024) uses experimental evidence on productivity gains from genAI usage and task-based predictions of genAI exposure. If we adjusted our estimate of labor cost savings from 33% down to the 27% from the experimental studies discussed in Acemoglu (2024), we would arrive at an aggregate productivity estimate of 0.9%. The remaining 0.2 percentage point gap is explained by different estimates of genAI adoption. Acemoglu (2024) uses expert assessments to estimate a wage-weighted share of all tasks for which genAI can be adopted of 4.6%. We use survey data on actual adoption to estimate a wage-weighted usage share of 5.7%, which yields slightly higher estimates.

Conclusion. Generative AI is a potentially important new technology. Its economic impact will ultimately be determined by how many people use it, how intensively they use it, and the productivity gains it provides. This paper documents results from the first nationally representative U.S. survey of genAI use at work and at home. Our data come from the Real-Time Population Survey (RPS), a survey that is constructed and weighted to be nationally representative and follows the same survey design as the CPS, a widely used national data source.

As of August and November of 2024, 39 percent of working-age adults reported using genAI, either for work or personal use. 9 percent of employed respondents used it every work day and 14 percent on some but not all workdays, bringing total weekly usage to 23 percent. ChatGPT is used most frequently, followed by Gemini, and our estimates of ChatGPT usage are similar to other recent surveys. We also validate our findings in a different survey panel, the Survey of Working Arrangements and Attitudes (SWAA).

We compare the speed of genAI adoption with PCs and the internet relative to the release date of the first mass market product for each technology. We find that overall adoption of genAI has been more rapid than these previous technologies. While work adoption rates were similar for genAI and PCs, adoption rates outside of work were much faster for genAI. Low adoption costs and the consumer focus of genAI products are a likely reason for relatively fast adoption rates, and the longer-run impacts of genAI remain uncertain. Still, it is notable that the speed of genAI adoption has been similar to the adoption of personal computers in the 1980s, since the latter led to an accelerate of productivity growth and rising inequality (e.g. Autor, Levy, and Murnane 2003).

We also find substantial variation in genAI adoption by education, age, gender, and occupation. genAI and computer adoption have similar demographic patterns, with greater early usage among younger, more educated workers in professional occupations. However, women were earlier users of PCs, in part because of rapid adoption of computers in white-collar clerical occupations.