'The Rapid Adoption of Generative AI' (Federal Reserve Bank of St Louis Working Paper Series, 2024) by Alexander Bick, Adam Blandin and David J Deming comments
Generative Artificial intelligence (AI) has rapidly emerged as a potentially transformative workplace technology. The large language model (LLM) ChatGPT debuted in November 2022, and by March 2024 the most common generative AI tools had been accessed more than three billion times by hundreds of millions of users each month (Liu and Wang, 2024). Several recent studies have found that generative AI 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). Yet other studies expect only modest impacts of AI on work, depending on how well AI substitutes for complex job tasks (Acemoglu, Autor, Hazell, and Restrepo, 2022; Bloom, Prettner, Saadaoui, and Veruete, 2024).
The ultimate impact of generative AI on the economy depends on how quickly and intensively the technology is adopted. Yet there is little systematic evidence of the extent to which generative AI is used at work and at home. Who uses generative AI, how much do they use it, and what do they use it for?
This paper presents results from the first nationally representative U.S. survey of generative AI adoption at work and at home. Our data come from the Real-Time Population Survey (RPS), a nationwide survey that asks the same core questions and follows the same timing and structure of the Current Population Survey (CPS), the monthly labor force survey conducted by the U.S. Census Bureau for the Bureau of Labor Statistics (BLS). We benchmark our survey to national estimates of employment and earnings, ensuring representativeness. Prior research has used the RPS methodology to study work from home during the COVID-19 pandemic, among other topics (Bick and Blandin, 2023; Bick, Blandin, and Mertens, 2023). The survey structure allows us to easily add and modify questions and to track generative AI usage over time within a large representative sample of the U.S. workforce.
We find that in August 2024, 39.4 percent of the U.S. population age 18-64 used generative AI, with 32.0 percent using it at least once during the week they were surveyed; 28.0 percent of employed respondents used generative AI at work, with most (24.2 percent) using it at least weekly; and 10.6 percent of the employed reporting daily usage at work. Generative AI use is more common outside of work, but less intensive. One in three respondents (32.7 percent) said that they used generative AI outside of work, but only 6.4 percent used it outside of work every day. ChatGPT is by far the most commonly used generative AI program, although many others are reported, including tools that embed AI inside standard office software packages (e.g., Microsoft Copilot).
How does the speed and intensity of the adoption of generative AI compare with other technologies? Prior research shows that better technologies are adopted faster, and the speed and intensity of technology adoption across countries is highly correlated with economic growth (Beaudry, Doms, and Lewis, 2010; Comin and Hobijn, 2010; Comin and Mestieri, 2018). We compare the speed of adoption of generative AI with two other technologies - the personal computer (PC) and the internet - using data from the CPS Computer and Internet Use Supplement and the International Telecommunication Union (ITU).
Generative AI has been adopted at a faster pace than PCs or the internet. Generative AI has a 39.5 percent adoption rate after two years, compared with 20 percent for the internet after two years and 20 percent for PCs after three years (the earliest we can measure it). This is driven by faster adoption of generative AI at home compared with the PC, likely because of differences in portability and cost. We find similar adoption rates at work for PCs and for generative AI. (Note that we cannot separate internet usage between home and work.)
Some scholars argue that generative AI could reduce workplace inequality (e.g., Autor 2024). However, similar to PC adoption, generative AI usage is more common among younger, more educated, and higher-income workers. This is notable because the PC revolution was followed by rising labor market inequality, with computers substituting for routine “middle-skill” tasks while complementing high-skilled labor (Autor, Levy, and Murnane, 2003). The one exception is gender. We find that men are 9 percentage points more likely to use generative AI at work and 7 percent more likely to use it at home. In contrast, PC adoption at work was more common for women, possibly because of the transition between typewriters and word processors and the high female share of secretaries and other administrative occupations.
Generative AI is used by workers in a broad range of occupations to perform many different workplace tasks. Generative AI adoption is most common in management, business, and computer occupations, with usage rates exceeding 40 percent. Still, one in five “blue collar” workers and one in five workers without a college degree use generative AI regularly on the job as well. This is consistent with Eloundou, Manning, Mishkin, and Rock (2024), who compare generative AI capabilities with the task content of work and find that many occupations will be affected. We asked workers whether they used generative AI to help them perform ten different job tasks, including writing, searching for information, interpreting data or text, coding, data analysis, and others. Among generative AI users at work, all ten tasks in our list had usage rates of at least 25 percent, with writing, interpreting, and administrative help ranked as the most helpful.
Using responses to questions about both the frequency and the intensity of work usage, we estimate that between 0.5 and 3.5 percent of all work hours in the U.S. are currently being assisted by generative AI. If we assume that generative AI increases task productivity by 25 percent - the median estimate across five randomized studies - this would translate to increase in labor productivity of between 0.125 and 0.875 percentage points at current levels of usage. However, this calculation assumes that small-scale studies are externally valid and should be treated with caution.
Our results line up broadly with other published surveys of generative AI usage. The most similar study to our is Humlum and Vestergaard (2024), who survey a representative sample of workers in eleven occupations in Denmark about their usage of ChatGPT at work. We find broadly similar usage rates in the occupations covered by both surveys, although the lack of clean correspondence between job codes across countries makes an exact comparison difficult.2 A Pew Research Center survey conducted in February 2024 found that 23 percent of adults in their survey reported ever having used ChatGPT, with higher rates of adoption for younger and more educated respondents (McClain, 2024). A Reuters online survey conducted in six countries in April 2024 found that 18 percent of U.S. respondents used ChatGPT at least weekly, compared to less than 10 percent in Argentina, Denmark, France, Japan, and the United Kingdom (Fletcher and Nielsen, 2024).
Our study shows that the generative AI is being adopted much faster than previous waves of AI technology. McElheran et al. (2024) find that less than 6 percent of firms had used frontier AI technologies such as machine learning, computer vision, and natural language processing in 2017. Similarly, Acemoglu, Autor, Hazell, and Restrepo (2022) find that only about 3 percent of U.S. firms had adopted predictive AI tools between 2016 and 2018 and Humlum and Meyer (2022) found similarly low adoption rates in Denmark in 2017. Generative AI may be adopted more rapidly because it targets consumers rather than firms. Bonney et al. (2024) report firm-level AI adoption using the Business Trends and Outlook Survey (BTOS), a Census Bureau study that asked firms about AI usage between December 2023 and February 2024. They found that AI adoption rose over the survey period from 3.7 percent in December to 5.4 percent in February, which is a rapid rise but still far below our estimates. Like Bonney et al. (2024), we also find that generative AI usage is higher in large firms. Still, gaps by firm size are far too small to explain the discrepancy between firm and worker usage, suggesting that workers are using generative AI even in firms that haven’t officially adopted it