The release of ChatGPT in November 2022 took the world by storm. Within months, it had accumulated over 100 million users and triggered the release of several other generative AI products, including Claude, Llama, Gemini, and Microsoft Copilot. It also spawned debate over the extent to which generative AI would transform our economy and society (e.g. Acemoglu et al. 2023, Bohren et al. 2024, Albanesi et al. 2023). On one side, many argued that generative AI would radically reshape our economy, delivering large gains in productivity but also labour market displacement. Others claimed that generative AI would remain a niche or novelty product, offering useful applications in a few jobs but only modestly impacting the broader economy. The stakes of this debate are high – for workers making career choices, for managers optimising their production process, for investors assessing the value of AI-related technologies, and for policymakers balancing concerns about growth and inequality.
The ultimate impact of generative AI on the economy hinges on the answers to two related questions. First, does the technology improve productivity in many important work tasks? Second, how quickly and intensively is generative AI being adopted by workers and firms?
An important obstacle to answering these questions is a lack of reliable, nationally representative data on generative AI adoption. In particular, we need to know how many people are using generative AI, which people are using it, how often are they using it, and for what tasks they are using it most.
To overcome this obstacle, we launched the first nationally representative survey on generative AI adoption at home and at work. Our data come from a new module in the Real-Time Population Survey (RPS), an online survey of working age adults in the US that has run since 2020 (Bick and Blandin 2023). Previously the RPS has been used to study the labour market impact of the Covid-19 pandemic (Bick and Blandin 2020) and the accompanying rise in working from home (Bick and Blandin 2023, Bick et al. 2024a). The RPS asks the same core questions as the Current Population Survey (CPS), the monthly labour force survey conducted by the US Census Bureau and the Bureau of Labor Statistics. This allows us to validate many of our measures against a trusted benchmark, but still leaves room for novel questions.
We first fielded our generative AI module in a smaller pilot study in June 2024. In this column, we focus on our results from our first full-scale survey two months later, in August 2024, which we discuss in much greater detail in our paper (Bick et al. 2024b).
How prevalent is generative AI adoption?
We find that in August 2024, 32% of the US population aged 18-64 used generative AI at least once in the week prior to the survey (see Figure 1). About 24% of employed respondents had used generative AI at work in the previous week, with 11% using it every workday that week. Generative AI use is more common outside of work, but less intensive: one in three respondents (32.7%) said they use generative AI outside of work, but only 6.4% had used it outside of work every day in the previous week.
Figure 1 Share of working age adults using generative AI
We asked respondents about which generative AI products they used. We found that ChatGPT was used most often, followed by Google Gemini. Our estimates closely align with other estimates of ChatGPT use. A survey conducted by the Pew Research Center in February 2024 found that 27% of US adults aged 18-64 reported ever having used ChatGPT (McClain 2024), compared to 28% in our survey six months later. A Reuters survey conducted in April 2024 found that 18% of US adults used ChatGPT at least weekly, compared to 19% of working age adults in our survey (Fletcher and Nielsen 2024). Unlike these previous studies, we asked about all generative AI use (not just ChatGPT); we asked separately about use at work and outside of work; and we collected more detailed data on which workers are using Generative AI, how frequently they are using it, and for what specific purposes.
How does the generative AI adoption compare to other technologies?
Should we view the pace of generative AI adoption thus far as fast or slow? One way to answer this question is to compare the speed of adoption of generative AI with the personal computer (PC) and the internet using data from the CPS Computer and Internet Use Supplement and the International Telecommunication Union (ITU).
We find that generative AI has been adopted at a faster pace than PCs or the internet (see Figure 2). Two years after the introduction of ChatGPT, 39% of individuals reported using generative AI (with 32% using it in the previous week), compared to 20% for PCs three years after the first mass computer was released (the IBM PC, in 1981) and 20% for the internet two years after it was opened for commercial traffic in 1995. The faster pace of adoption by generative AI is driven by faster adoption outside of work compared with the PC, likely because of differences in portability and cost.
Figure 2 The trajectory of computer, internet, and AI adoption
By contrast, we find that the pace of generative AI adoption at work is similar to that of PCs. (We cannot separate internet usage between home and work.) We also find that generative AI adoption patterns by age, education, and weekly earnings percentile are strikingly similar to early PC adoption patterns. In other words, early generative AI adoption looks very similar to PCs in terms of how many people are using the technology at work and which kinds of workers are using it. This is potentially relevant for forecasting the eventual economic impact of generative AI. For example, the introduction of computers and related information technologies was associated with a large increase in earnings inequality and, with a lag, large productivity gains.
How are workers using generative AI?
To investigate how generative AI is impacting work, we asked workers how they used it. Respondents were presented with the list of tasks shown in Figure 3 and asked to select those for which they used generative AI at work in the previous week.
Figure 3 In which specific tasks is AI most useful?
The most common tasks were those that involved writing (“Writing Communications”) or information collection/analysis (“Searching for Facts or Information”, “Documentation or Detailed Instructions”, “Interpreting/Translating/Summarizing”). However, maybe the most important takeaway is that generative AI is used for a wide range of tasks, with usage rates at or exceeding 25% for all ten defined tasks. When we asked respondents to rank the tasks for which they used generative AI in order of how helpful the technology was in completing the task, eight of the ten tasks were ranked in the top two by at least 10% of respondents.
Looking ahead: How much could generative AI increase labour productivity?
Finally, we examined how intensely respondents employed generative AI on days that they reported using it. This allowed us to estimate the share of all work hours in the US that involve generative AI. We estimate that between 1% and 8% of all work hours in the US are currently assisted by generative AI. If generative AI increased productivity by 25% – the median estimate across five different randomised studies – generative AI could plausibly grow labour productivity by between 0.2% and 2.1% at current levels of usage (Brynjolfsson et al. 2023, Cui et al. 2024, Dell’Acqua et al. 2023, Noy and Zhang 2023, Peng et al. 2023).
To summarise, the impact of generative AI on the US economy hinges on how many workers use the technology and how much of their work is impacted by the technology. Thus far, generative AI has been adopted quite rapidly compared with other transformative technologies. We also find that workers are using generative AI for a wide range of tasks. At the same time, generative AI is used directly in less than 10% of work hours, so it is still far from ubiquitous. Whether generative AI represents a truly transformational technology will therefore depend on whether it continues to spread throughout the labour market. We plan to release a new survey every few months so that we can closely monitor these developments.
References
Acemoglu, D, D Autor and S Johnson (2023), “How AI can become pro-worker”, VoxEU.org, 4 October.
Albanesi, S, A da Silva, J Jimeno,A Lamo and A Wabitsch (2023), “Artificial intelligence and jobs: Evidence from Europe”, VoxEU.org, 29 July.
Bick, A and A Blandin (2020), “Real-time labour market estimates during the 2020 coronavirus outbreak”, VoxEU.org, 6 May.
Bick, A and A Blandin (2023), “Employer Reallocation During the COVID-19 Pandemic: Validation and Application of a Do-It-Yourself CPS”, Review of Economic Dynamics 49: 58-76.
Bick, A, A Blandin, A Caplan and T Caplan (2024a), “Measuring Trends in Work From Home: Evidence from Six U.S. Datasets”, CEPR Discussion Paper 19495.
Bick, A, A Blandin and D Deming (2024b), “The Rapid Adoption of Generative AI”, CEPR Discussion Paper 19515.
Bick, A, A Blandin and K Mertens (2023), “Work from Home Before and After the COVID-19 Outbreak”, American Economic Journal: Macroeconomics 15(4): 1-39.
Bohren, N, R Hakimov and R Lalive (2024), “Creative and Strategic Capabilities of Generative AI: Evidence from Large-Scale Experiments”, CEPR Discussion Paper 19507.
Brynjolfsson, E, D Li and L Raymond (2023), “Generative AI at work,” Working Paper.
Cui, Z, M Demirer, S Jaffe, L Musolff, S Peng and T Salz (2024), “The Effects of Generative AI on High Skilled Work: Evidence from Three Field Experiments with Software Developers”, Working Paper.
Dell’Acqua, F, E McFowland, E Mollick, H Lifshitz-Assaf, K Kellogg, S Rajendran, L Krayer, F Candelon and K Lakhani (2023), “Navigating the jagged technological frontier: Field experimental evidence of the effects of AI on knowledge worker productivity and quality,” Working Paper.
Fletcher, R and R Nielson (2024), “What does the public in six countries think of generative AI in news?”, in press.
McClain, C (2024), “Americans’ use of ChatGPT is ticking up, but few trust its election information”, in press.
Noy, S and W Zhang (2023), “The productivity effects of generative artificial intelligence”, VoxEU.org, 7 June.
Peng, S, E Kalliamvakou, P Cihon and M Demirer (2023), “The impact of AI on developer productivity: Evidence from github copilot,” Working Paper.
