Key Points
We estimate that 40 percent of current GDP could be substantially affected by generative AI. Occupations around the 80th percentile of earnings are the most exposed, with around half of their work susceptible to automation by AI, on average. The highest-earning occupations are less exposed, and the lowest-earning occupations are the least exposed.
AI’s boost to productivity growth is strongest in the early 2030s, with a peak annual contribution of 0.2 percentage points in 2032. After adoption saturates, growth reverts to trend. Because sectors that are more exposed to AI have faster trend TFP growth, sectoral shifts during the AI transition add a lasting 0.04 percentage point boost to aggregate growth.
Compounded, TFP and GDP levels are 1.5% higher by 2035, nearly 3% by 2055, and 3.7% by 2075, meaning that AI leads to a permanent increase in the level of economic activity.
Caution is required in interpreting these projections of AI’s impact, which are based on limited data on AI’s initial effects. Future data and developments in AI technology could lead to a significant change in these estimates.
In ongoing work, PWBM is estimating the impact of AI on the federal budget. In very preliminary analysis, we estimate that AI could reduce deficits by $400 billion over the ten-year budget window between 2026 and 2035.
The Projected Impact of Generative AI on Future Productivity Growth
Generative artificial intelligence (AI) technologies such as large language models (LLMs) are increasingly being used to perform tasks that rely on digital tools and information processing. To gauge the implications for productivity growth, we combine the task-based framework developed by Acemoglu (2024) with a projected adoption timeline informed by the historical diffusion of comparable technologies like the commercial web and cloud computing services. We also consider an expanded range and intensity of occupations impacted by generative AI.
Acemoglu’s (2024) framework considers how AI will affect the different tasks performed by workers.1 Generative AI tools may increase worker efficiency for some tasks, replace workers altogether by automating others, or cause the introduction of new tasks. Acemoglu shows that ultimately the impact of AI on total factor productivity (TFP) growth depends on two factors:
The share of economic activity impacted by AI tools.2
The cost savings from adopting AI tools.
We estimate these values based on recent studies of AI’s applicability to different tasks and outcomes for businesses that adopted generative AI tools.
We define a job as exposed if at least 50 percent of the activities performed in that job could be automated by generative AI.3 Drawing on Eloundou et al.’s (2024) detailed classification of which tasks can be partly or fully automated with AI tools, we estimate that around 42 percent of current jobs are potentially exposed to AI. Table 1 shows our estimates of exposure to AI automation by type of occupation, based on SOC major groups. To support additional research on this topic, a supplementary data file available for download provides exposure estimates for 784 detailed occupational categories that are aggregated in Table 1.
| Occupation Group | Exposure to AI Automation (% of tasks) |
|---|---|
| Office and Administrative Support Occupations | 75.5 |
| Business and Financial Operations Occupations | 68.4 |
| Computer and Mathematical Occupations | 62.6 |
| Sales and Related Occupations | 60.1 |
| Management Occupations | 49.9 |
| Legal Occupations | 47.5 |
| Arts, Design, Entertainment, Sports, and Media Occupations | 45.8 |
| Architecture and Engineering Occupations | 40.7 |
| Life, Physical, and Social Science Occupations | 31.0 |
| Educational Instruction and Library Occupations | 29.5 |
| Community and Social Service Occupations | 27.5 |
| Healthcare Practitioners and Technical Occupations | 23.1 |
| Protective Service Occupations | 20.7 |
| Transportation and Material Moving Occupations | 20.0 |
| Food Preparation and Serving Related Occupations | 18.1 |
| Personal Care and Service Occupations | 17.5 |
| Healthcare Support Occupations | 15.5 |
| Production Occupations | 14.4 |
| Installation, Maintenance, and Repair Occupations | 13.1 |
| Farming, Fishing, and Forestry Occupations | 9.7 |
| Construction and Extraction Occupations | 8.9 |
| Building and Grounds Cleaning and Maintenance Occupations | 2.6 |
Source: PWBM based on estimates from Eloundou et al.’s (2024) and data from the Bureau of Labor Statistics.
Figure 1 shows the distribution of employment by automation potential based on the categorization of tasks provided by Eloundou et al. (2024). For 29 percent of jobs, there is no potential to substitute AI for workers. For another 29 percent, AI could automate less than half of the activities required. Only around 1 percent of jobs are completely exposed to automation, meaning they could be performed entirely by AI without significant oversight from a human. However, for more than a quarter of U.S. employment, AI could perform between 90 and 99 percent of the work required with minimal oversight.
Source: PWBM based on estimates from Eloundou et al.’s (2024) and data from the Bureau of Labor Statistics.
Note: Estimates are based on employment data for 2024.
To estimate the corresponding share of economic activity, each exposed job must be weighted by its earnings and by the percentage of the job that could be done by AI. Figure 2 plots the relationship between wages and exposure to automation at the occupation level. The points show the average share of tasks that can be performed by AI in each percentile of occupational wage distribution. The solid line shows the underlying relationship between AI exposure and occupational earnings at the detailed occupation level.
Occupations at the bottom of the wage distribution are the least exposed to AI, since many of these jobs are predominantly manual labor or personal services. Exposure generally rises with earnings until the 80th-90th percentiles, which include programmers, engineers, and other professionals. In these relatively high-wage occupations, around half of the work could be performed by generative AI on average. This proportion falls sharply in the highest-earning occupations, which include business executives, athletes, and medical specialists.4
Source: PWBM based on estimates from Eloundou et al.’s (2024) and data from the Bureau of Labor Statistics.
Notes: The points show the average exposure to AI by percentile of the occupational annual wage distribution. The solid line is the kernel density of exposure to AI by annual wage percentile for 784 detailed occupations, scaled to the binned percentiles. Within occupations, tasks are considered as having zero exposure if less than 50 percent of the task components could be performed by AI. Estimates are based on data for 2021 to 2024 to account for volatility in annual wages.
Combining occupation-level employment, wages, and exposure, we estimate that 40 percent of current labor income is potentially exposed to automation by generative AI. This value is almost twice the value estimated by Acemoglu (2024); the differences are explained in the Appendix.5 Following Acemoglu, we make two further assumptions: First, we assume the share of GDP exposed to AI is the same as the share of labor income. Second, we project that it will not be feasible and profitable to actually adopt AI tools for all exposed tasks. Based on the findings of Svanberg et al. (2024), we assume that 23 percent of exposed tasks will eventually be automated.
Putting it all together, we estimate that just under 10 percent of current GDP is likely to be impacted over time. We project this share will grow to around 15 percent over the next two decades as sectors that are more exposed grow faster than the rest of the economy and the proportion of tasks that can be profitably automated increases.
Based on studies of real-world generative AI applications, we assume labor cost savings of roughly 25 percent on average from adopting current AI tools. Table 2 lists several studies of AI adoption and summarizes their results. These studies find gains ranging from around 10 to 55 percent, with an average of around 25 percent. We project the average labor cost savings will grow from 25 to 40 percent over the coming decades.
| Study | Domain | Outcome |
|---|---|---|
| Brynjolfsson et al. (2023) | Customer service with a generative AI assistant. | 14% increase in task completion rate. |
| Jabarian and Henkel (2025) | Job interviews with a generative AI voice agent. | 17% increase in job starts; 18% increase in retention rate. |
| Noy and Zhang (2023) | Basic professional writing with ChatGPT-3.5. | 40% increase in speed; 18% increase in output quality. |
| Peng et al. (2023) | JavaScript programming with GitHub Copilot. | 56% increase in speed. |
| Cui et al. (2025) | Software development with GitHub Copilot. | 26% increase in task completion rate. |
| Wiles et al.(2023) | Job applications with algorithmic resume writing assistance. | 8% increase in likelihood of hire. |
| Dell'Acqua et al. (2023) | Management consulting with GPT-4 (experimental setting) | 12% increase in task completion rate; 25% increase in speed. |
Sources: See links in the first column.
AI’s impact on TFP depends on how quickly productivity-enhancing tools are actually adopted. Data on the diffusion of generative AI tools remains limited, but Bick et al. (2025) find that 26.4 percent of workers used generative AI at work in the second half of 2024, while 33.7 percent of adults used it outside of work. Using data from the Real-Time Population Survey (RPS), they show that early adoption patterns of AI for work are broadly similar to the adoption of personal computers (PCs) in the early 1980s.6
Figure 3 compares these early AI adoption rates against historical diffusion measures for previous mass-market technologies: the PC, the internet, smartphones, and cloud computing. The first panel shows adoption at work and the second panel shows adoption outside of work. The horizontal axis shows the numbers of years since the launch of each technology’s first mass-market product – in generative AI’s case, the launch of OpenAI’s ChatGPT in 2022.
(a) At Work
(b) Outside of Work
Sources: Real-Time Population Survey, Current Population Survey, Annual Business Survey, Pew Research Center, and PWBM based on Kalyani et al. (2024).
Note: See the Appendix for the definitions of mass-market introduction and descriptions of each series.
Adoption of major new technologies in the workplace follows a remarkably consistent pattern. By the end of the first decade after technology’s introduction, between 40 and 50 percent of workers are using it. Adoption slows sharply but continues in the following decade. Historical experience of technology adoption outside of work is more varied but shows similar patterns. In both cases, the figures for generative AI use in 2024 from the RPS suggest somewhat faster adoption than previous technologies.
Source: PWBM based on estimates from Eloundou et al.’s (2024) and data from the Bureau of Labor Statistics.
Figure 4 presents suggestive evidence that AI adoption is already affecting the labor market. Calculating the change in employment since 2021 by level of exposure to AI (based on the Eloundou et al. classification), we find that job growth has stagnated in occupations with most AI automation potential. For jobs that can be performed entirely by generative AI, employment fell sharply in 2024 and was 0.75 percent lower than in 2021 (however, recall from Figure 1 that these jobs make up only around 1 percent of total employment). In occupations with high AI exposure (90 to 99 percent of tasks can be automated) the shift has been less dramatic, but employment growth has slowed significantly since 2022.7
Combining estimates of exposure, cost savings, and adoption, we project generative AI’s contribution to TFP growth over the next several decades. Figure 5 plots our projections. Despite examples of successful AI adoption such as those described in Table 2, we estimate that AI’s impact on TFP growth remains small today - 0.01 percentage points (pp) in 2025 - as most businesses have yet to deploy and gain experience with AI tools. Over the next decade, AI’s contribution will grow for three reasons:
Generative AI tools will increasingly be applied to tasks exposed to AI productivity gains.
AI technologies will improve, increasing the potential cost savings from applying AI to a given task.
The share of economic activity exposed to AI will rise due to long-running sectoral trends, with sectors relatively more exposed to AI (such as software development and professional services) growing faster than the rest of the economy.
We project that AI will boost TFP growth by 0.09pp in 2027, 0.18pp in 2030, and peak in the early 2030s at around 0.2pp. As new adoption slows in the 2030s due to declining remaining opportunities to employ additional AI tools productively, the impact on TFP growth diminishes to around 0.1pp by end of the decade and continues to shrink thereafter. We project that TFP growth will be persistently higher by a little less than 0.04pp even after adoption saturates and TFP growth returns to trend. This occurs because sectors that were more exposed to AI also have faster trend TFP growth, and those sectors will make up a larger share of the economy as a result of AI-driven productivity gains.
Source: PWBM
Because these are growth effects, their cumulative impact – the impact on TFP levels – is what matters for living standards. Cumulating projected growth contributions implies that the level of TFP will be around 1.5% higher by 2035, 3% higher by 2055, and 3.7% higher by 2075 relative to a noAI path. Put differently, AI makes the economy permanently larger, but once adoption saturates, the ongoing growth rate itself returns to trend—aside from the small, persistent lift from sectors that benefit more from AI.
The hump-shaped pattern in Figure 2 mirrors past digital technology diffusion waves: an initial acceleration as diffusion proceeds, then normalization as the new technology becomes ubiquitous. The ultimate scale of the economic gain is anchored by (i) how much activity is truly exposed and (ii) how large the tasklevel cost savings become. Our parameterization implies a material but not transformative macroeconomic effect under current evidence.
Given the early nature of generative AI, considerable caution is required when interpreting these projections. Currently, our analysis does not account for:
AI-driven changes in product quality, including potential decreases in product quality.
The emergence of new products and labor tasks as a result of AI adoption.
The potential impact of AI tools on innovation, which could feed back into TFP growth in either a positive or negative direction.
Our projections will be updated as more data and more information about these issues becomes available.
Following Acemoglu’s (2024) derivation, we project AI’s long run impact on TFP using the following relationship:
Long run percentage change in TFP = GDP share of tasks impacted by AI × Percentage cost savings from AI
Each of the two terms on the right hand side can be further broken down:
GDP share of tasks impacted by AI = GDP share of tasks exposed to AI × Share of exposed tasks worth automating
Percentage cost savings from AI = Percentage labor cost savings from AI × Labor’s share of income from exposed tasks
We add a third term to Acemoglu’s expression to account for the adoption timeline:
Percentage change in TFP = Long run percentage change in TFP × Generative AI adoption rate
The adoption rate is defined relative to the universe of tasks that can be profitably automated using generative AI.
We estimate exposure to generative AI at the SOC 6-digit occupation level based on the automation exposure metric developed by Eloundou et al.’s (2024). This metric classifies tasks into one of five “T” categories based on what percentage of the task an AI could complete at high quality:8
T0: 0 percent
T1: 0-50 percent
T2: 50-90 percent
T3: 90-99 percent
T4: 100 percent
To estimate the average exposure within each occupation, we take the following steps:
Define tasks in T0 and T1 (0 percent and 0-50 percent automation potential, respectively) as not exposed to generative AI.
Assign a single numeric exposure value to each of the other three categories, equal to the midpoint of that category’s range: 70 percent for T2 (50-90 percent exposed), 95 percent for T3 (90-99 percent exposed), and 100 percent for T4.
Aggregate from tasks to occupations by taking the average exposure value across all tasks performed in each occupation.
We obtain data on employment and annual wages from the Bureau of Labor Statistics’ Occupational Employment and Wage Statistics (OEWS). For a small number of occupations with irregular work schedules, OEWS does report an annual wage, only an hourly wage. In these cases, we use the American Community Survey to estimate the average number of hours worked per year.
To obtain exposure estimates for the U.S. economy, we aggregate occupations based on total wage income in 2021 to 2024. We use multiple years of data to account for volatility in annual wages.
Though we follow the same general approach as Acemoglu (2024), our estimate of the share of GDP exposed to AI is 40 percent relative to his 20 percent. The difference comes from how we interpret Eloundou at al.’s automation exposure metric.
As described above, we translate Eloundou at al.’s five categories into a numeric exposure value based on their original definitions in terms of the percentage of work that can be performed by AI. However, Eloundou at al. present another version of the metric in which they assign a numeric score to each category to facilitate comparisons with other quantities. As shown in Table A1, they assign values of 0, 0.25, 0.5, 0.75, and 1 to categories T0 through T4, respectively. This assignment treats each category as an equal increment of 0.25, rather than using the percentages from the original category definitions to quantify them.
| Automation Category (Task Exposure %) | Eloundou et al. (2024) | Acemoglu (2024) | PWBM Baseline | PWBM Expanded |
|---|---|---|---|---|
| T0 (0%) | 0 | 0 | 0 | 0 |
| T1 (0-50%) | 0.25 | 0 | 0 | 0.25 |
| T2 (50-90%) | 0.5 | 0 | 0.7 | 0.7 |
| T3 (90-99%) | 0.75 | 0.75 | 0.95 | 0.95 |
| T4 (100%) | 1 | 1 | 1 | 1 |
Acemoglu interprets these scores as reflecting the percentage of work in each category that can be automated by AI and replaces them with values of 0 percent, 25 percent, 50 percent, 75 percent, and 100 percent. As in our approach, he considers a task to be exposed to AI if more than 50 percent of its components could be performed by generative AI. However, he interprets a score of 0.5 on the Eloundou at al. metric (category T2) as a task for which AI could do 50 percent of the work – rather than 50 to 90 percent as in Eloundou at al.’s original definition – and so excludes the T2 category from his definition of exposed tasks. Acemoglu’s measure of exposed tasks is therefore limited to categories T3 and T4, which contain tasks that could be at least 90 percent performed by AI, according to Eloundou at al.
To combine tasks into occupations, Acemoglu weights tasks in each “T” category by its score value. Since he only includes categories T3 and T4, this amounts to giving tasks in T3 a weight of 75 percent relative to tasks in T4. Similarly, we aggregate tasks according their percentage automation potential, but assign a weight of 70 percent to tasks in T2 and 95 percent to tasks in T3, corresponding to the midpoints of the range of automation potential contained within each category.
Hence, Acemoglu’s lower estimate of the GDP share exposed to AI reflects two major differences: 1) We consider a task to be exposed if AI could perform 50-90 percent of the task components, while Acemoglu treats these tasks as not exposed to AI; 2) We assign greater weight to exposed tasks in category T3, where AI could perform 90-99 percent of task components (we assign a weight of 95 percent vs. Acemoglu’s 75 percent).
Finally, note that if we instead use Eloundou at al.’s original scores directly as weights, we obtain a GDP-weighted share of exposed tasks of about 38 percent – essentially the same as our 40 percent. Alternatively, if we expand our baseline approach and include tasks in category T1 with a weight of 0.25 as shown in the “PWBM Expanded” column of Table A1, our estimate rises to 48 percent (category T1 means AI can perform some of the task but less than 50 percent). This broader definition of exposure to AI implies a somewhat larger estimate of AI’s impact on TFP: an increase of 4.4 percent by 2075, compared with 3.7 percent in our baseline case.
The sources for the technology adoption series plotted in Figure 3 are as follows:
PCs and Internet: Current Population Survey, Computer and Internet Use Supplement (CPS CIU) – The CPS is a household survey that has periodically included questions about computer and internet use. Households are asked whether they have a computer at home or have the internet at home. Workers are asked whether they use a computer at work or use the internet at work. We report these shares directly in Figure 3.
Cloud computing, at work: Annual Business Survey (ABS) – The ABS is a survey of businesses that periodically includes questions about the use of certain technologies. Businesses are asked whether they use cloud computing as part of their production process or method. We report the share of total employment that is in firms indicating they use cloud computing in Figure 3.
Smartphone, at work: PWBM estimates based on Kalyani et al.’s (2024) dataset – We define “adoption” at the 4-digit NAICS industry level: an industry is considered to have adopted a technology if at least 2 percent of job postings in that industry include a reference to the technology.9 The adoption rate we report in Figure 3 is the number of industries that have crossed this threshold, weighted by total job postings in that industry.
Smartphone, outside of work: Pew Research Center – Pew surveys U.S. adults about whether they own their own smartphone. We report the share who do own a smartphone directly in Figure 3.
Generative AI: Real-Time Population Survey (RPS) – The RPS is an online survey of households designed to be consistent with the CPS. Beginning in 2024, it included questions about use of generative AI modeled after the CPS CIU questions on PCs and the internet. We report the results directly in Figure 3.
For each technology, we define “mass-market introduction” as the launch of the first major mass-market product based on the technology. The products and corresponding years we identify for each technology are as follows:
PCs: The IBM PC, released in 1981.
Internet: Netscape, which became the first widespread web browser and made an initial public offering in 1995.
Cloud computing: Amazon S3, launched in 2006.
Smartphone: The iPhone, released in 2007.
Generative AI: ChatGPT, released in 2022.
To project adoption of generative AI at work, we combine the adoption curves for the other four technologies, extract a common time component, and extrapolate the curve out to 35 years from mass-market introduction. Following the historical adoption pattern, we assume that generative AI will be almost fully incorporated into production within the next decade and a half, but that some new adoption continues at a gradual and diminishing pace over the following decades.
This analysis was produced by Alex Arnon under the direction of Kent Smetters. Vidisha Chowdhury provided research assistance. Mariko Paulson prepared the brief for the website. We thank Heidi Williams for feedback on an earlier version of this brief.
Media (only): For the fastest response, email us at inquiries-pwbm@wharton.upenn.edu.
All other responses: Please use our Contact Us.
Sign up for PWBM Breaking News, Alerts and Newsletter. Unsubscribe anytime.
-
Throughout this brief, we use the general term “AI” for convenience. However, our analysis is limited to generative AI tools such as LLMs. ↩
-
Specifically, the GDP share of tasks impacted by AI. ↩
-
Automation in this context includes processes that replace workers with generative AI tools and those that augment workers by enabling them to work more efficiently. ↩
-
This profile is strikingly similar to the pattern of actual AI adoption by earnings identified by Bick et al. (2025). ↩
-
Acemoglu estimates that the exposed share of labor income is 20 percent. See the appendix for a discussion of the differences between the two estimates. ↩
-
The RPS is a proprietary online survey designed to be consistent with the official federal survey that provides data on unemployment and other aspects of the labor market. ↩
-
See also Brynjolfsson et al. (2025), who find that employment is falling for the most exposed workers. ↩
-
See Eloundou et al.’s supplementary materials. The exact bounds of each category are not clearly identified. ↩
-
We select a threshold of 2 percent in order to match adoption estimates based on the Kalyani et al. job postings data with adoption reported in the 2019 Annual Business Survey (ABS) for technologies included in both datasets. The ABS does not report adoption of smartphones generally, so our estimate for smartphones is not directly calibrated to any external source. ↩
category_label fill_color sort_order emp_pct 0% not_exposed 0 29.35434322 0-50% not_exposed 1 28.60504372 50-90% ai_exposed 2 15.44754568 90-99% ai_exposed 3 25.48053792 100% ai_exposed 4 1.112529457
wage_bin auto_score_pct kde_auto_score_pct 0 21.75732377 7.705803571 1 19.27801724 8.626883079 2 15.94551215 9.553711893 3 9.182159326 10.48138234 4 33.49041525 11.40500558 5 17.50723623 12.31978509 6 20.95517677 13.22108566 7 22.6985834 14.10449643 8 13.93983957 14.96588659 9 29.29715049 15.80145267 10 17.34012213 16.60775704 11 24.09131196 17.38175716 12 23.57677045 18.12082603 13 10.71693396 18.82276414 14 18.02753913 19.48580375 15 15.79551489 20.10860652 16 16.42430124 20.69025553 17 47.28753092 21.23024294 18 37.66347854 21.72845444 19 27.99986306 22.18515155 20 20.79443093 22.60095284 21 14.14218588 22.9768147 22 27.40364583 23.31401251 23 35.87155388 23.61412237 24 11.8006993 23.87900364 25 28.55185905 24.1107823 26 27.28824544 24.31183487 27 14.49174922 24.48477255 28 27.00491033 24.63242499 29 42.10082498 24.75782337 30 26.46078585 24.86418202 31 33.12289026 24.95487809 32 25.05364305 25.03342887 33 20.899462 25.10346623 34 28.34234078 25.16870807 35 10.32008468 25.23292655 36 18.89354402 25.2999132 37 26.23605517 25.37344117 38 27.63499949 25.45722493 39 16.72500201 25.55487803 40 22.95483984 25.66986959 41 31.27267574 25.80548026 42 31.86216375 25.9647586 43 23.67865896 26.15047879 44 16.54446248 26.36510053 45 24.60694206 26.61073217 46 7.705803571 26.88909783 47 21.14912497 27.20150924 48 12.07924583 27.54884314 49 25.57778055 27.93152434 50 22.86858974 28.34951524 51 38.76365976 28.80231145 52 13.96519544 29.28894383 53 20.28468407 29.8079865 54 31.04761905 30.35757046 55 23.01333042 30.93540218 56 34.07353992 31.53878655 57 24.86696269 32.16465321 58 22.84255113 32.80958541 59 43.06876796 33.46985052 60 30.55822709 34.14143105 61 25.44352291 34.8200555 62 36.84048972 35.50122806 63 33.54115085 36.18025655 64 31.97923753 36.85227809 65 26.22559524 37.51228217 66 40.75172956 38.15513112 67 31.29199967 38.77557813 68 57.49979589 39.3682833 69 34.20954385 39.92782846 70 19.87207927 40.44873172 71 41.39281426 40.92546308 72 52.16596832 41.35246218 73 44.05546449 41.72415998 74 51.80119959 42.03500566 75 38.84152035 42.27950019 76 47.37934121 42.45223784 77 34.82921807 42.54795662 78 32.47405438 42.5615982 79 41.26059843 42.48837744 80 45.67643309 42.32386122 81 40.21001077 42.06405533 82 38.21145812 41.70549807 83 39.63584807 41.24535806 84 48.96516331 40.68153349 85 47.20828564 40.01274954 86 46.78777604 39.23865018 87 36.94833563 38.35988037 88 47.36327569 37.37815475 89 48.09683698 36.29630881 90 39.42307407 35.11832917 91 41.7208476 33.84935986 92 48.0744354 32.49568232 93 51.96757285 31.06466788 94 43.26077084 29.56470213 95 46.19388191 28.00508207 96 36.44584295 26.39588765 97 27.03157117 24.74783065 98 22.07924212 23.07208468 99 19.91839492 21.38010096
panel years_since_start label value at work 3 PC (1984-2003) 24.6 at work 8 PC (1984-2003) 36.8 at work 12 PC (1984-2003) 45.8 at work 16 PC (1984-2003) 49.8 at work 20 PC (1984-2003) 54 at work 22 PC (1984-2003) 56.1 at work 2 Internet (1997-2003) 16.6 at work 6 Internet (1997-2003) 38.9 at work 8 Internet (1997-2003) 42.3 at work 0 Smartphone (2007-2019) 8.9614877 at work 3 Smartphone (2007-2019) 20.01965261 at work 4 Smartphone (2007-2019) 23.55322086 at work 5 Smartphone (2007-2019) 26.72824969 at work 6 Smartphone (2007-2019) 29.2552313 at work 7 Smartphone (2007-2019) 31.34105866 at work 8 Smartphone (2007-2019) 33.08338801 at work 9 Smartphone (2007-2019) 35.30182979 at work 10 Smartphone (2007-2019) 37.58953234 at work 11 Smartphone (2007-2019) 40.01818889 at work 12 Smartphone (2007-2019) 42.58417606 at work 12 Cloud computing (2018-2022) 52.8 at work 16 Cloud computing (2018-2022) 54.2 at work 2 Generative AI (2024) 26.4 outside of work 3 PC (1984-2012) 8.2 outside of work 8 PC (1984-2012) 15 outside of work 12 PC (1984-2012) 22.9 outside of work 16 PC (1984-2012) 36.6 outside of work 19 PC (1984-2012) 51 outside of work 20 PC (1984-2012) 56.3 outside of work 22 PC (1984-2012) 61.8 outside of work 26 PC (1984-2012) 69.7 outside of work 28 PC (1984-2012) 74.1 outside of work 29 PC (1984-2012) 76.7 outside of work 30 PC (1984-2012) 75.6 outside of work 31 PC (1984-2012) 78.9 outside of work 2 Internet (1997-2012) 18 outside of work 5 Internet (1997-2012) 41.5 outside of work 6 Internet (1997-2012) 50.4 outside of work 8 Internet (1997-2012) 54.7 outside of work 12 Internet (1997-2012) 61.7 outside of work 14 Internet (1997-2012) 68.7 outside of work 15 Internet (1997-2012) 71.1 outside of work 16 Internet (1997-2012) 71.7 outside of work 17 Internet (1997-2012) 74.8 outside of work 4 Smartphone (2011-2024) 35 outside of work 5 Smartphone (2011-2024) 44.125 outside of work 6 Smartphone (2011-2024) 54.28571429 outside of work 7 Smartphone (2011-2024) 56.75 outside of work 8 Smartphone (2011-2024) 68 outside of work 9 Smartphone (2011-2024) 73 outside of work 11 Smartphone (2011-2024) 77 outside of work 12 Smartphone (2011-2024) 81 outside of work 14 Smartphone (2011-2024) 85 outside of work 16 Smartphone (2011-2024) 90 outside of work 17 Smartphone (2011-2024) 91 outside of work 2 Generative AI (2024) 33.7
YEAR category_label sort_order pct_change 2021 0% 0 0 2022 0% 0 5.468385107 2023 0% 0 8.325515968 2024 0% 0 9.898975767 2021 0-50% 1 0 2022 0-50% 1 5.290119879 2023 0-50% 1 8.864128848 2024 0-50% 1 11.53608508 2021 50-90% 2 0 2022 50-90% 2 5.830227695 2023 50-90% 2 9.537026893 2024 50-90% 2 11.88876313 2021 90-99% 3 0 2022 90-99% 3 4.190694715 2023 90-99% 3 6.088199643 2024 90-99% 3 6.532440466 2021 100% 4 0 2022 100% 4 1.932929508 2023 100% 4 2.44898515 2024 100% 4 -0.734282219
year value 2025 0.010802786 2026 0.047540123 2027 0.090757899 2028 0.118744222 2029 0.15371562 2030 0.179360364 2031 0.193256977 2032 0.196060982 2033 0.192257829 2034 0.184990589 2035 0.174517689 2036 0.161059678 2037 0.144697917 2038 0.128171601 2039 0.110082595 2040 0.091724666 2041 0.075640517 2042 0.062992858 2043 0.054890094 2044 0.050377588 2045 0.048035456 2046 0.046539723 2047 0.045455927 2048 0.045260591 2049 0.04512266 2050 0.044412418 2051 0.043268051 2052 0.042093501 2053 0.041928716 2054 0.041810551 2055 0.040957377 2056 0.039671456 2057 0.038339403 2058 0.038149917 2059 0.038347493 2060 0.038526317
