Key Information
- adoption rate:39.4 of the American population aged 18 to 64 use generative AI,28% of the employees use it at work, and nearly 1/9 use it every day. Apart from work, 32.7% of the interviewees use it, but only 6.4% use it every day.
- Adopted speed: the adoption speed of generative AI is faster than that of personal computers and the internet. The adoption rate in two years is 39.5%, while that of the latter is only 20% in the same period.
- Features: generative AI is the most common in management, business and computer professions, with a usage rate of over 40%; Nearly 1/4 of blue-collar workers also use it. The usage scenario covers multiple tasks such as writing, information retrieval, and data analysis, and the usage rate exceeds 25%.
-
Economic impact: the contribution of generative AI to working hours is estimated to be 0.5% to 3.5%. Assuming that productivity increases by 25%, the overall labor productivity may increase by 0.125 to 0.875 age points. There are obvious differences in utilization rate among different industries, with the highest utilization rate in finance/insurance/real estate industry (51%) and the lowest utilization rate in leisure/accommodation industry (15%). Out-of-work use is more common but less intensive.
----------
summary
generative artificial intelligence (GenAI) is a potentially important new technology, but its impact on the economy depends on the speed and intensity of adoption. This paper reports the results of the first national representative survey in the United States, involving the adoption of generative AI in work and family in the United States. In August 2024, 39 percent of the US population aged 18 to 64 used generative AI. In the week before the survey, more than 24 percent of employees used it at least once, and almost one in nine people used it at work every day. Historical data show that the adoption speed of generative AI has exceeded the popularity of personal computers and the Internet. Generative AI is a general technology widely used in various occupations and work tasks, whether in the workplace or at home.
1 Introduction
the rapid rise of generative artificial intelligence (GenAI) has become a technology that may transform the workplace. Large language model (LLM)ChatGPT was introduced in November 2022. By March 2024, the most commonly used generative AI tool had been accessed by hundreds of millions of users for more than 3 billion times (Liu and Wang,2024). Many recent studies have found that generative AI improves employee 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). However, other studies predict that the impact of AI on work is only moderate, which depends on the alternative ability of AI in complex work tasks (Acemoglu, Autor, Hazell and Restrepo,2022;Bloom, Prettner, saadaoui and Veruete,2024).
The ultimate impact of generative AI on the economy depends on the speed and intensity of the technology. However, there is currently a lack of systematic evidence to demonstrate the extent to which generative AI is used at work and at home. Who is using generative AI, and what are their frequencies and uses?
This paper presents the results of the first national representative survey in the United States, involving the adoption of generative AI in work and family in the United States. Our data comes from the real-time population survey (RPS), a nationwide survey that uses the monthly Labor Force Survey (CPS) conducted by the U.S. Census Bureau for the Bureau of Labor Statistics (BLS). Same core issues and time structure. We compare the findings with national employment and income estimates to ensure representativeness. Previous studies have used the RPS method to study topics such as home work during COVID-19 epidemic (Bick and Blandin,2023;Bick, Blandin and Mertens,2023). The survey structure enables us to easily add and modify problems and track the use of generative AI in a large representative sample of the us workforce.
We found that in August 2024, 39.4 percent of the American population aged 18 to 64 used generative AI, and 32.0 percent of them used AI at least once within the survey week; 28.0% of the employees interviewed use generative AI at work, and most of them (24.2%) use it at least once a week; 10.6% of the employees report that they use generative AI at work every day. Generative AI is more commonly used outside of work, but its strength is relatively low. One third of the interviewees (32.7%) said they used generative AI outside of work, but only 6.4% used it every day outside of work. ChatGPT is obviously the most commonly used generative AI program, although many other tools have been mentioned, including AI tools embedded in standard office software packages (for example, Microsoft Copilot).
How is the speed and intensity of generative AI compared with other technologies? Previous studies have shown that better technologies are adopted faster, and the speed and intensity of technology adoption in various countries are highly correlated with economic growth (Beaudry, Doms and Lewis,2010;Comin and Hobijn,2010;Comin and Mestieri,2018). We will compare the adoption speed of generative AI with personal computer (PC) and Internet technologies, and use CPS computers and supplementary surveys on Internet use, as well as data from the International Telecommunication Union (ITU).
The adoption speed of generative AI is faster than that of personal computers and the Internet. In two years, the adoption rate of generative AI was 39.5%, while that of Internet was 20%, the three-year adoption rate of personal computers is also 20% (this is the earliest data we can measure). This phenomenon is mainly due to the faster adoption of generative AI in homes than personal computers, which may be related to differences in portability and cost. We found that PC and generative AI have similar adoption rates at work. (Please note that we cannot distinguish the use of the Internet between home and work.)
some scholars believe that generative AI may reduce workplace inequality (for example, Autor 2024). However, similar to the adoption of personal computers, the use of generative AI is more common among young, higher education and higher income workers. This point is worth noting, because after the personal computer revolution, the inequality in the labor market rose. Computers replaced regular "medium-skilled" tasks and supplemented high-skilled labor (Autor, Levy and Murnane,2003). The only exception is gender. We found that men are 9 percent age points more likely to use generative AI at work than women and 7 percent age points higher in families. In contrast, the adoption rate of personal computers at work is more common among women, which may be due to the transition from typewriters to word processors and the high proportion of women in secretaries and other administrative professions.
Generative AI is widely used in various professions to perform various tasks. The adoption of generative AI is the most common in management, business and computer professions, with a usage rate of over 40%. Nevertheless, every five "blue collar" workers and every five workers without university degrees also regularly use generative AI at work. This is consistent with the research of Eloundou, Manning, Mishkin and Rock(2024). They compared the ability of generative AI with the task content of work and found that many occupations would be affected. We asked workers whether to use generative AI to help them perform ten different tasks, including writing, information retrieval, data or text interpretation, coding, and data analysis. Among the users who use generative AI at work, the usage rate of the ten tasks we listed is at least 25%, of which writing, explanation and administrative support are considered to be the most helpful.
According to the answers to the frequency and intensity of work use, we estimate that 0.5% to 3.5% of the working time in the United States is assisted by generative AI. If we assume that generative AI will increase task productivity by 25%-this is the median estimate in five random studies-this will translate into an increase in labor productivity, ranging from 0.125 to 0.875 percentage points. However, this calculation assumes that small-scale studies are effective externally and should be treated with caution.
Our results are basically consistent with other published generative AI usage surveys. The most similar studies to ours are Humlum and Vestergaard(2024), who surveyed representative sample workers in 11 occupations in Denmark to understand their use of ChatGPT at work. We found that the utilization rates of occupations covered in the two surveys were roughly similar, although accurate comparisons became difficult due to the lack of clear correspondence between work codes in various countries. A survey conducted by the Pew Research Center in February 2024 found that 23% of the adults interviewed said they had used ChatGPT, and the adoption rate of young and highly educated respondents was higher (McClain,2024). An online survey conducted by Reuters in six countries in April 24found that 18% of American respondents used ChatGPT at least once a week, while in Argentina, Denmark, France, the proportion of Japan and Britain is less than 10%(Fletcher and Nielsen,2024).
Our research shows that the adoption speed of generative AI far exceeds the previous wave of AI technology. McElheran et al. (2024) found that in 2017, less than 6% of companies used cutting-edge AI technologies, such as machine learning, computer vision and natural language processing. Similarly, Acemoglu, Autor, Hazell and Restrepo(2022) found that between 2016 and 2018, only about 3% of American companies adopted predictive AI tools, while Humlum and Meyer(2022) in 2017, Denmark's adoption rate was also low.
The reason why generative AI can be adopted more quickly is that it is mainly for consumers, not enterprises. Bonney et al. (2024) used the business trend and prospect survey (BTOS) to report the AI usage at the company level. The survey inquired about the AI usage of enterprises between december 2023 and february 2024. They found that during the investigation, the AI adoption rate rose from 3.7% in December to 5.4% in February. Although this increase was rapid, it was still far lower than our estimate. Consistent with Bonney et al. (2024), we also found that generative AI is more commonly used in large enterprises. However, the gap between enterprise sizes is too small to explain the difference between the use of companies and employees, which indicates that even in companies that have not formally adopted generative AI, employees are also using this technology.
The structure of this paper is as follows: the second part describes the investigation methods and data, the third part presents our main results, and the fourth part is the conclusion.
2 data sources and measurements
2.1 Real-time population survey (RPS)
our data source is the real-time population survey (RPS), a national labor market survey for adults aged 18 to 64 in the United States (for detailed discussions, see Bick and Blandin,2023). RPS is conducted online through Qualtrics, a large commercial survey provider, and has collected multiple rounds of survey data every year since 2020. RPS is designed to mirror the Current Population Survey (CPS) in key dimensions. The RPS survey adopted the same word-for-word wording on the population characteristics and labor market results of the basic CPS and CPS outflow groups, and copied the complex problem sequence needed to obtain the labor market results, to ensure consistency with CPS (U.S. Census Bureau, 2015). Copying key parts of existing high-quality surveys ensures the comparability of survey concepts, enabling researchers to verify RPS results with a wide range of benchmarks with larger sample sizes, and construct sample weights when necessary.
However, RPS also collects information not included in CPS. The new problems in RPS have been used to study the relationship between employee reallocation between companies, home work, interstate migration and inflation and job hunting (Bick and Blandin,2023;Bick, blandin and Mertens,2023;Bick, Blandin, Mertens and Rubinton,2024;Pilossoph and Ryngaert,2023). In june and august 2024, RPS introduced a module to measure the use of generative AI at work and at home.
RPS generated very similar statistical data in terms of employment, working hours, income, industry composition and employee tenure during the epidemic (Bick and Blandin,2023). This is partly attributed to the detailed and iterative implementation of the investigation to match the question wording and other details of CPS.
2.2 Sample
interviewees on The Qualtrics panel recruit online and can participate in the survey to obtain 30% to 50% of the fees charged by Qualtrics as compensation (we pay $6.90 for each completed survey). The Qualtrics panel has about 15 million members and is not a random sample of the US population. However, researchers can instruct Qualtrics to target the Survey Invitation to specific population groups. The RPS sample aims to be nationally representative in a wide range of population characteristics, including gender, age, race and nationality, education, marital status, number of family children, census area and household income in the past 12 months.
We conducted a pilot survey in June 2024 and received 2551 replies. Subsequently, a complete investigation was launched in August 2024 and 5014 replies were received. Both investigations were conducted within the same week as CPS conducted corresponding investigations. We removed 14 and 33 respondents from the June and August surveys respectively because the industries and/or occupations they reported were military. Another 9 interviewees were excluded from each survey because they reported that they were employed but also called housewives, retired or unemployed.
The survey in august 2024 made many improvements on the basis of the pilot survey. This is a larger sample and contains questions about the usage intensity of generative AI and the usage span across work tasks. We have also improved the coding of industries and occupations and made other minor improvements. We found that the results of overlapping problems in the two surveys were very similar, with a slightly higher work usage in August, although this increase was not statistically significant. Therefore, for the sake of simplification, this article only reports the results of August. Appendix B copied the main charts of this article using the survey data of June 2024.
Table 1: Composition of C PS and RPS samples in August 2024
remarks: the first column reports the samples of variables targeted by Qualtrics in the sampling process in the pre-population survey (CPS) in august 24. Employment status is the only untargeted variable. The second column reports the sample composition of the real-time population survey (RPS) in August 24. Samples from both datasets are limited to the civilian population aged 18 to 64. The third and fourth columns report the same results of employees (who were on duty and absent last week).
The first two columns in table 1 compare the population composition of CPS and RPS under the target sampling procedure in our main survey in august 2024. The most significant difference is that individuals aged 18 to 24 and whose educational background does not exceed that of high school graduates are underrepresented in RPS, while individuals with annual income of $50,000 or less are overrepresented in RPS. The lower half of Table 1 compares the employment status in CPS and RPS, and these statistical data are not set in the sampling process. According to the definition of CPS, individuals classified as unemployed are overrepresented in RPS.
The third and fourth columns in Table 1 compare the demographic composition of employed interviewees in CPS and RPS. This has generally improved the balance, although there are still some differences.
2.3 sample weight and verification
to resolve these differences, we use the iterative scale adjustment (raking) algorithm of Deming and Stephan(1940) to construct the sample weight. Our application of raking algorithm ensures that the weighted sample proportion of key population characteristics matches the proportion in CPS. We also used more detailed classifications than the educational and marital status included in Qualtrics sampling targets and interacted all of these classifications with gender. In addition, our sampling weight replicates the classification of employment status, whether overall or based on a series of our target characteristics, to ensure that some variables have a large enough sample size. We also included occupations in the weighted scheme. This requires us to eliminate the other 108 and 112 observation values respectively due to the lack of professional codes in the June and August surveys. Appendix c.1. provides detailed information on the categories targeted by our weighted scheme.
Figure 1: verification check
remarks: the figure on the left uses unweighted RPS data, and the figure on the right uses weighted RPS data. Both figures use RPS sample respondents. All graphs use weighted CPS data. The weekly income data shows the samples of employees aged 18 to 64 in R PS and CPS-ORG in August 24. the weekly income is lower than the highest code of CPS by 3960 USD, and the implied hourly wage is at least $7.25 of the federal minimum wage. The sample sizes of RPS and CPS are 2184 and 6078 respectively. Occupational data samples are employed interviewees aged 18 to 64 in R PS and CPS in August 2024, with sample sizes of 3216 and 42987 respectively.
Bick and Blandin(2023) show that RPS copies CPS well in many dimensions that are not set in the sampling procedure or are not included in the weighted scheme, including regular and actual weekly working hours, proportion of workers paid by hour, weekly income distribution, industry composition and working term. The panels (a) and (B) of fig. 1 compare the regular weekly income distribution of RPS and CPS, without weighting and weighting respectively. The unweighted distribution is similar, and the weighting further improves the fitting degree.
The panels (c) and (d) of Fig. 1 compare the occupation shares in RPS and CPS, respectively, without weighting and weighting. The two samples are basically the same without weighting. The correlation between the two samples is 0.87, and the management profession is especially excessively represented in RPS. For this reason, we apply career weights to all further analyses. Panel (d) shows that this adjustment mechanically balances the sample professionally.
2.4 Measurement of the use of generative AI
define and guide issues. The generative AI module first gives the definition of generative AI: Generative AI is an artificial intelligence that generates text, images, audio, or video in response to prompts. Some examples of generative AI include ChatGPT, Gemini, and Midjourney. As generative AI is a relatively emerging technology, we think it is very important to provide conceptual definitions and some specific examples. We avoid mentioning specific generative AI methods such as "large language models" because these terms are too technical for the general audience, and we hope our definition can cover a wide range of methods. At the same time, we mentioned some popular examples of generative AI products, because we think some interviewees may be more familiar with these product names than the broader concept of generative AI.
After defining generative AI, the module asks respondents if they have heard of the concept before the survey. Respondents who answer "no" will skip the rest of the module, while respondents who answer "yes" will continue to answer the next question in the AI module.
Generate AI usage at work. For employed interviewees, the next question is whether to use generative AI at work: Do you use generative AI at work? (No/Yes) this problem is intended to correspond to similar problems in the supplementary investigation of CPS computer and internet use. For detailed discussion, see section 2.5. Respondents who answer "no" will skip subsequent questions about job-related generative AI. Respondents who answer "yes" will be asked additional questions about the use of generative AI at work, which are divided into two categories. The first type is related to the intensity used by generative AI. We asked them about the number of days they used generative AI in the past week and the average daily usage time in these days. The second type of questions asks the interviewees what specific products they use, what help generative AI provides in specific tasks, and some broader questions about the use and benefits of generative AI.
The use of generative AI outside of work. The last part of the generative AI module asks about usage outside work: Do you use generative AI outside work? (No/yes) respondents who answer "no" will jump directly to the end of the generative AI module. Respondents who answer "yes" will be asked a set of additional questions similar to the "at work" section.
2.5 measurement of computer and internet use
since 1984, CPS has carried out an occasional supplementary investigation involving computer and internet use, called supplementary investigation on computer and internet use (CIU). Issues related to our research were investigated by CIU in 1984, 1989, 1993, 1997, 2001, 2003, 2007 and 2009. All respondents who accepted basic CPS questions also received CIU questions.
We are concerned about two sets of questions related to computer use in the CIU supplementary survey. The first question is about the use of computers at work: Do you use computers [directly] at work? (No/Yes) The second question asks about the use of computers in your home: Do you use computers [directly] at home? (No/Yes) CIU asks about "home" computer usage, while we ask about "out of work" generative AI usage. This means that our wording is not exactly the same as the CPS problem, but our broader wording has two advantages. First of all, home work is not uncommon nowadays. Asking about the use of generative AI in the home will not clearly distinguish between work and non-work use. Secondly, many people access the generated AI through mobile devices, asking for home use alone may not be able to capture non-work use outside the home.
We also use a question about Internet use in CIU: Do you use the internet anywhere? (No/Yes) this question was asked in CIU's 2001, 2003, 2007 and 2009 surveys. Different from computer-related questions, it is not limited by location.
In order to calculate the internet popularity before 2001, we used data provided by the International Telecommunication Union (ITU). ITU cooperated with the World Bank to collect data on Internet use in the United States and other countries since 1995. Based on the Internet usage data of national regulatory agencies and service providers, they estimated the proportion of Internet access in the population based on the number of subscribers (Pen ++ a-L 'opez et al., 2009).
3 results
figure 2 shows our main results. The first bar chart shows that 39.4% of RPS respondents in August 2024 said they used generative AI at work or at home. About 32 percent of the interviewees reported using AI at least once a week before the survey, while 10.6 percent reported using AI every day last week. About 28% of the employed interviewees used generative AI during their work in August 2024, of which the vast majority (24.1%) used it at least once in the past week and 10.9% used it every day. Out-of-work use is more common (32.7%), but the use intensity is slightly lower, 25.9% used at least once last week, 6.4% used every day. Figure A.1 of the appendix shows the proportion of respondents using specific generative AI products. ChatGPT has the highest usage (28.5%), followed by Google Gemini(16.3%).
Figure 2: proportion of working-age adults using generative AI
remarks: this figure shows the proportion of interviewees who use AI at work, outside work, and overall (outside work or work). The usage intensity is divided into daily use last week (dark blue), use at least one day a week but not daily use (medium blue), and unused last week (light blue). The data source is RPS in August 2024, and the age range is 18 to 64 years old. "Work sample" is employed individual (N = 3216); Other bar charts include all interviewees (N = 4682).
Figure 3: different population characteristics used by AI at work
remarks: the figure shows the proportion of interviewees using AI in work by sex, age, education and major. The usage intensity is divided into daily use last week (dark blue), use at least one day a week but not daily use (medium blue), and unused last week (light blue). The data source is RPS in August 2024, and the age range is 18 to 64 years old. The sample of this figure is employed individuals (N = 3216). The samples of university majors are employed individuals with bachelor's degree or above. STEM majors include biology, agriculture, environment, physics and related sciences; Computer, mathematics and statistics; And engineering. "Business/communication/economy" includes business, communication and economics. "Humanities/others" includes all other majors.
Figure 3 shows the adoption rate of AI generation in work by sex, age, education and major. The survey shows that 32% of men use generative AI at work, while 23% of women use AI. The use of generative AI decreased with age, from about 34% among workers under 40 years old to 17% among workers over 50 years old. About 40% of workers with bachelor's degree or higher use generative AI, while about 20% of workers without university degree. 46% of workers majoring in science, technology, engineering or mathematics (STEM) use generative AI at work, while 40% of workers majoring in business, economics or communication, other majors (including humanities and social sciences) accounted for 22%. Figure A.2 of the appendix shows the usage of generation AI outside work by population characteristics. The overall trend is similar, but the difference is not as obvious as that used in work.
Appendix table A.1 shows the adoption of generative AI at work and the multiple regression coefficients of population characteristics and occupations. We found that the pattern described here is generally valid under multiple specifications, so for simplicity, we focus on simple bivariate comparison.
3.1 Comparison between the adoption of generative AI and personal computer and Internet
figure 4 compares the adoption speed of generative AI with other two Technologies (personal computer and internet). Horizontal axis measurement relative to the adoption of the first mass market product. The first mass market computer was IBM PC, which was released in August 1981 and sold more than one million units (Abbate, 1999). This means that the first data point we obtained from CIU was three years after the popularity of the public. The first AI model that sells at least one million subscriptions is ChatGPT, which was released in November 2022, only two years before our survey. Finally, we set the mass market availability date of the internet as april 1995, when the national science foundation (NSF) closed NSFnet and allowed the internet to carry commercial traffic (Leiner et al., 2009). This is also the year of Netscape's initial public offering (IPO).
Figure 4: trajectory of computer, internet and AI
remarks: the figure shows the usage of three technologies at work, including AI, computers and the Internet. The horizontal axis indicates the year since the launch of the first mass market product of each technology. We regard 2022 as the year of AI release, that is, the year of ChatGPT release. 1995 was the year when the Internet was launched, because NSF shut down NSFNet and allowed the Internet to transmit commercial traffic. 1981 was the year when computers were launched, and IBM PC was released exactly in that year. AI data comes from RPS (solid blue dot) in August 2024. Computer data comes from the supplementary survey of computer and internet use of CPS from 1984 to 2003 (hollow red square). We have drawn two estimates of Internet use: CPS computer and internet use supplementary survey (dark green triangle) and International Telecommunication Union (ITU)(cyan triangle) from 2001 to 2009. RPS and CPS samples were all individuals aged 18 to 64. The RPS sample size is N = 4682. ITU samples are individuals of all ages.
To facilitate comparison, we measure the comprehensive use of these three technologies, because we cannot distinguish between working and non-working Internet use. The blue dots in figure 4 repeat the 39.4% adoption rate of the AI generation in figure 2. The red square describes the adoption of personal computers from the third year to the 22nd year, covering the CPS supplementary survey from 1984 to 2003. The adoption rate of personal computers rose steadily, from 20% in the third year to nearly 70% in the 22nd year. Dark green triangle shows the Internet usage during the 6th to 14th years, covering the CPS survey from 2001 to 2009. At that time, problems about Internet usage were added. By the 6th year, the adoption rate of the Internet had been about 60%. Light green triangle presents the adoption of the US internet collected by the International Telecommunication Union (ITU) from 0 to 26 (for example, from 1995 to 2021). The Internet adoption rate rapidly increased from 20% in the second year to 60% in the seventh year, and then steadily increased to 90% in the next 20 years. The data of the two datasets in overlapping years is very close.
Figure 4 shows that, so far, the adoption of generative AI is faster than that of personal computers and the Internet. The reason why generative AI is adopted faster than personal computers is that it is more widely used outside of work and may be related to differences in portability and cost. Appendix figure A.3 only compares the adoption of generated AI and personal computers at work, using CIU data (we cannot distinguish between internet use at home and at work). We found that the adoption rate of generative AI in the second year was 28%, while that of personal computers in the third year was 25%.
In terms of education and income, generative AI is similar to the early adoption mode of personal computers. Appendix figure A.4 shows that after three years of popularization in the mass market, about 42% of workers with bachelor's degree or above use personal computers at work, however, only 20% of workers have not obtained university degrees. For generative AI, we found that the adoption rate of educational level is almost the same. Appendix figure A.5 shows the usage of generative AI and personal computers based on the percentile of weekly income. The usage patterns of the two technologies are very similar. When the income rises to about 85th percentile, the usage rate increases and then decreases slightly.
Although the adoption patterns of the two technologies are generally very similar, the only exception is gender. Appendix figure A.6 shows that 32% of men use generative AI at work, while only 23% of women use AI. In contrast, by 1984, only 22% of men used personal computers at work, compared with 30% of women. One possible explanation is that the proportion of women in offices and administrative support positions is relatively high, while the adoption rate of personal computers is close to 50%.
3.2 Use of generative AI in various occupations and tasks
figure 5a shows the adoption of generative AI by occupation. We collected the job names of the interviewees through free text responses, and then matched them to the Standard Occupational Classification (SOC) code using analytic algorithms. The recognition success rate reached 97%.
Figure 5: AI usage by occupation and industry group
remarks: figure 5a shows the proportion of interviewees using AI in work by occupation. The personal service profession combines SOC code 31-39, including medical support, protection services, food preparation and services, cleaning and maintenance, and personal care. Blue collar profession combines SOC code 4753, including construction, mining, installation, maintenance and repair, production, transportation and transportation. The usage intensity is divided into daily use last week (dark blue), use at least one day a week but not daily use (medium blue), and unused last week (light blue). The data source is RPS in August 2024, and the age range is 18 to 64 years old. The sample of this figure is employed individuals (N = 3191). Figure 5b shows the proportion of respondents using AI in work by industry. The usage intensity is also divided into daily use last week (dark blue), use at least one day a week but not daily use (medium blue), and unused last week (light blue). The data source is RPS in August 2024, and the age range is 18 to 64 years old. The sample of this figure is employed individuals (N = 3216).
At work, the AI adoption rate of computer/Mathematics and management professions is the highest, about 49%. The utilization rate of business, finance and education was also high, 42% and 38% respectively. However, the adoption of generative AI in many professions is relatively common. In addition to personal services, at least 20% of workers in all major occupational groups use generative AI at work. Interestingly, in the "blue collar" occupation (including construction and mining, installation and maintenance, skilled production and transportation and transportation), 22% of workers use generative AI at work.
Figure A.7 of the appendix shows the adoption of generation AI outside work by occupation of interviewees. The overall ranking is roughly the same, although the usage rate is higher, ranging from 27% in personal services to 47% in management. Figure 5b shows the results by industry. Workers in the finance, insurance and real estate industries have the highest usage rate of generative AI (51%), while those in the leisure and accommodation industries have the lowest usage rate (15%). Appendix figure A.8 compares the adoption rates of generative AI and personal computers by occupation. The use of personal computers is more concentrated in a few occupations, from more than 90% of computer and mathematics occupations to less than 10% of most blue-collar and personal service occupations.
Figure 6: Which specific tasks are AI most useful?
Remarks: this figure shows what AI users are most helpful in reporting AI tasks. Panel (a) involves tasks at work; Panel (B) involves tasks outside work. In addition to work and work, the interviewees first provided a list of tasks and were asked to select the tasks completed with AI assistance last week. Subsequently, respondents were asked to rank these Selected tasks according to the help level of AI in completing tasks. This figure reports the proportion of AI users who rank a specific task as the first (AI is the most helpful in this task) or the second. Since some interviewees chose fewer than two tasks, the sum of the histogram is unnatural. The data source is RPS in August 2024, and the age range is 18 to 64 years old. The samples of panels (a) and (B) are employed individuals (N = 3216) and all individuals (N = 4682), respectively.
We also asked RPS interviewees about which tasks generate AI is most useful. Among the interviewees who have used generative AI in the past week, we showed the list of tasks in figure 6 and asked them to select tasks that have used generative AI help in the past week. They can also fill in other tasks. Subsequently, the interviewees were asked to sort the selected tasks according to the help level of generative AI in completing the tasks.
Figure 6 reports the proportion of top two tasks to respondents. In the work, the highest ranking tasks are writing (38%), administrative tasks (27%) and interpretation/translation/summary of text or data (23%). Overall, the ranking is relatively uniform. At least 10% of the respondents in the ten tasks in the list listed the eight tasks as the top two. Apart from work, the highest ranking tasks are writing (27%), explanation/Translation/summary (23%) and personal assistance (21%). Similar to work usage, at least 10% of the respondents in the 11 tasks in the list listed the eight tasks as the top two.
Appendix figure A.9 reports the proportion of respondents using generative AI in each task. The overall order is roughly the same as Figure 6. The usage rate of all ten tasks at work is at least 25%, of which the most common tasks are writing (57%), information search (49%), and obtaining detailed instructions (48%). Apart from work, the most common tasks are personal assistance (such as lists, schedules, etc.), ideas and tips for creative projects, and writing. In general, generative AI is widely used in jobs and tasks.
3.3 The potential of generative AI to improve labor productivity
in addition to asking the interviewees how often they use generative AI, we also pay attention to its daily use intensity. Specifically, we asked the interviewees whether they used 15 minutes or less, 15 minutes to an hour, or more than an hour on working days and rest days using generative AI. Appendix figure A.10 reports the intensity of daily use of generative AI at work and at home. In general, 25% of AI users use more than one hour at work, and 52% of users use 15 to 60 minutes during use. There was a positive correlation between use intensity and frequency, with 42% of daily users reporting using one hour or more per day. The usage mode is similar beyond work.
We use the use intensity and frequency of generative AI to estimate the proportion of all working hours assisted by generative AI. For each interviewee, we multiply the working hours reported by the upper and lower limits of the frequency and intensity of the self-reported generative AI. For example, suppose a worker used generative AI on some workdays last week and used it between 15 and 60 minutes per day. For the lower limit, we assume that the interviewee spent 15 minutes on a certain day. The upper limit is to assume that they use generative AI for 59 minutes every day except for one day. For those respondents who reported using generative AI but did not use it last week, we assume that the usage time is zero.
By adding these data of all interviewees, it is concluded that the lower limit and upper limit of the weekly working hours assisted by generative AI are 0.5% and 3.5% respectively. This overall proportion is relatively small, mainly because a large number of workers reported that they did not use generative AI, reaching 76%.
Given these data, to what extent can generative AI improve labor productivity? Five recent studies have estimated the impact of generative AI on task productivity through experimental or quasi-experimental design. Noy and Zhang(2023) paid college-educated professionals for writing tasks and found that ChatGPT increased productivity by 40%. Brynjolfsson, Li and Raymond(2023) found that dialogue assistants based on generative AI increased productivity by 14%. Dell'Acqua et al. (2023) found that in pre-registration tasks that they think are suitable for AI assistance, the productivity of providing strategic consultants with access permissions for generative AI has increased by about 25%. Cui et al. (2024) found that in a large electronic manufacturing company, GitHub Copilot, an AI-based coding assistant, was randomly provided for software developers, increasing its productivity by 26%. Peng, Kalliamvakou, Cihon and Demirer(2023) studied the influence of GitHub Copilot on a single coding task and found that its productivity increased by 56%.
If we (slightly arbitrarily) multiply the median estimate of about 25% by the ratio of working time assisted by the generative AI, we estimate that it may increase labor productivity by 0.125 to 0.875% at the current level of use. If the adoption speed of generative AI remains the same as that of past technologies such as personal computers and the internet, and the use intensity or task composition remain unchanged, these figures will double in the next decade. However, we need to warn that this calculation is highly speculative. Workers and enterprises may first use generative AI in their most productive applications, which means that with the expansion of use, revenue decline may occur. On the other hand, with the passage of time, technology may become more advanced and applicable.
4 Conclusion
Generative AI has rapidly emerged as an important new technology, and its impact on the economy depends to a large extent on the speed and intensity of adoption. This paper reports the results of the first national representative survey in the United States on the use of generative AI at work and at home. Our data comes from the real-time population survey (RPS), which is constructed and weighted to ensure national representativeness and follows the same survey design as the widely used national data source CPS. We found that in August 2024, 39.4 percent of people aged 18 to 64 in the United States reported using generative AI, 28.0 percent of whom were used at work and nearly 1/9 of workers reported using AI daily.
We compare the adoption speed and intensity of generative AI with two revolutionary technologies-personal computers and the Internet. We found that the adoption of generative AI is faster than these two technologies. We also found that generative AI is widely used in a variety of occupations and work tasks. Nearly half of computer, mathematics and management workers are using generative AI, while nearly 1/4 of blue-collar workers are also using AI. We asked the interviewees about the practicability of generative AI in 11 different tasks, such as writing, administrative support, interpretation and summary of text or data, and coding. Our survey shows that the usage rate at work exceeds 25% in all ten tasks. In general, we found that generative AI is indeed a common technology (Eloundou, Manning, Mishkin and Rock,2024).
We estimate that about 0.5 to 3.5% of the working hours in the United States are supported by generative AI. Assuming that the productivity improvement of recent experimental studies is effective externally, this indicates that generative AI may increase labor productivity by 0.125 to 0.875 percent age points at the current use level, but we need to warn, this calculation should be considered highly speculative because it is based on specific assumptions.
Our research results provide many directions for future research. In particular, with the maturity of technology, it is very important to track the adoption of generative AI and monitor its wide use in workers, enterprises and occupations. Future RPS surveys will include more detailed questions about the frequency and intensity of generative AI adoption so that we can track its impact on the evolving US economy.
references: Bick, Alexander, Blandin, Adam, and Deming, David J. "The Rapid Adoption of beautai." NBER Working Paper No. 66, September 2024. National Bureau of Economic Research. https://www.nber.org/papers/w32966
---[This article is finished]] ---
this article is reprinted from Andy730 public account