The Rapid Adoption of Generative AI

New evidence from a nationally representative survey

Welcome back! Today we take a break from our weekly series about the history of technology-fueled labor market disruption to talk about some new research.

This morning my coauthors and I released the results of the first nationally representative survey of generative AI usage at work and at home. You can find a copy of the NBER working paper here, ungated version here, and here is a Harvard Project on Workforce blog post about it.

---

(Sidebar – many have asked why I chose the name Forked Lightning. It comes from Dylan Thomas – “Though wise men at their end know dark is right / Because their words had forked no lightning they / Do not go gentle into that good night”. I take it as a nudge for ivory tower denizens like me to try and fork some lightning out there in the real world…)

---

Generative AI burst onto the scene with OpenAI’s release of ChatGPT in November 2022. Lots of reporting suggests that even people inside OpenAI were shocked about the rapid ascent in popularity of their new product. According to Web traffic and Google Trends data, the top 40 generative AI tools (including ChatGPT) receive nearly three billion monthly visits from users all around the world.

A recent Pew survey found that 23% of adults said they had used ChatGPT, with usage at work more than doubling between March 2023 (8 percent) and February 2024 (20 percent).¹ The Pew survey has some limitations - it was not nationally representative, it did not cover other generative AI products, and it did not really dig into how the technology is being used. Still, it suggests surprisingly high usage for such a new, unproven technology.

Generative AI is everywhere, but I encounter a lot of skepticism of it in my professional circles.² Most people I know have used it, but few have integrated it into their daily routines. They seem to view it more as a curiosity. Adding to their skepticism, surveys of AI usage by companies suggest very low adoption rates. Recent work using a well-established Census Bureau survey of businesses found that only 5.4% of firms were using generative AI, up from 3.7% in December.

New evidence on generative AI usage

There’s a lot of hype around generative AI. We wanted to know – who is using it, how much are they using it, and what are they using it for?

To answer these questions, we needed a high-quality, nationally representative survey. The gold standard for information about the labor market is the Current Population Survey (CPS), a nationally representative monthly survey of about 60,000 U.S. households that has been running for more than 80 years. The CPS is run by the Census Bureau on behalf of the Bureau of Labor Statistics (BLS), a division of the U.S. Department of Labor.³ The CPS produces monthly estimates of the unemployment rate and other key statistics that track the health of the labor market.

In the past, the CPS has added supplementary surveys on technology adoption. In 1984 it began asking questions about personal computer (PC) usage, and in 2001 it added some questions about the internet. Because it was the main credible source of data on early technology adoption, lots of important papers were written using data from the CPS Computer and Internet Use (CIU) supplement.

Since we don’t (yet?) have a CPS supplement on generative AI use, my coauthors and I decided to create one.

During the pandemic, Alex Bick and Adam Blandin created the Real-Time Population Survey (RPS) to ask questions about remote work and other timely topics. The RPS is replicates the sampling frame, question wording, ordering and survey structure, and timing of the CPS. Since we collect the same data in the same way at the same time, we can benchmark our estimates of employment, earnings, and other characteristics in the RPS to the CPS. They match very closely, and we use survey weights to make them match exactly.

Think of our survey as a clone of the CPS, but with questions added about generative AI usage. (You can dig into the Appendix materials if you want to verify that it worked.)

Ok, enough preamble. What did we find?

In August 2024, 39.4% of adults ages 18-64 said they use generative AI, either at work or at home. 32 percent said they used it at least one day in the week prior to being surveyed, and nearly one in nine (10.9 percent) said they used it every day last week. 28% used it at work, 24.2% used it at least once in the last week, and 10.6% used it every day. Usage at home is higher than usage at work (32.7%), but less intensive, with only 6.4% saying they used it every day outside of work.⁴

This is way higher than I expected! When I told friends and colleagues about our survey, I would always ask them to guess first. The 39.4% figure is higher than most people’s guesses (except for the college students, who tend to guess that *almost* everybody is using it).

Here are three reasons why I think we should believe the numbers. First, the Pew survey I mentioned above was of all adults, including retirees. If you restrict their numbers to ages 18-64, usage increases to 27%. Second, Pew only asked about ChatGPT, whereas we asked about all programs (we gave a list and allowed them to select all that they had used). ChatGPT was the most common program (28.5%), but it was still less than half of the total, with others like Google Gemini (16.3%) taking significant market share. Accounting for other numbers should push the 27% figure from Pew even higher. Third, we know that generative AI usage is rising rapidly, and the Pew survey was conducted 6 months ago.

We also looked at demographic differences in generative AI adoption. Usage is more common among young and highly educated workers, and more common among men compared to women.

How does generative AI stack up to other technologies?

As I mentioned above, the CPS has in the past asked about computer and internet usage in their CIU supplement.⁵ We replicated the same question structure as the CIU so we could compare the speed of adoption of generative AI to PCs and the internet.

We measure “year zero” of adoption based on the mass-market availability of each technology. For generative AI that’s the 2022 release of ChatGPT, the first LLM to sell more than a million subscriptions. For PCs that’s the 1981 release of the IBM PC, the first computer to sell more than a million units. For the internet, we choose 1995, the year that the NSF opened it up to commercial traffic (it’s also the year of the Netscape IPO).

Our data show that generative AI adoption has been faster (39.4% after 2 years) than PCs and the Internet (20% and 30% after three years respectively). For PCs and the internet, usage more than doubled between year 3 and year 15, which if the trend holds implies that generative AI usage would exceed 80% by 2036.

Why has generative AI adoption happened so quickly, and what does that say about its future impact? One important thing to remember is that fast adoption was only possible because nearly everyone has access to a computer and to the internet. These technologies build on each other.

In some sense, you can think of PCs and the Internet as the hardware and software “base layers” that allow complementary innovations like generative AI to be quickly and cheaply distributed. As I said a couple of weeks ago, I expect that to be true of generative AI as well. We are still learning about the new possibilities unlocked by the availability of this general form of intelligence. MS Office, web browsers, and Google search were killer apps for PCs and the Internet, and we are still waiting for the killer app of generative AI.

GPTs are GPTs

The other major surprise to me in our survey was the incredible breadth of generative AI use across occupations and job tasks. I expected to find high usage rates in computer, STEM, management and business jobs. And we did! Nearly half of workers in those occupations are using generative AI, and more than one in six are using it every day.

However, I did not expect to find that 22 percent of blue-collar workers are using generative AI at work, or that usage rates were 20 percent or higher for all major occupation categories except personal services. (Blue collar is construction, installation and maintenance, precision production, transportation, and manual labor jobs.)

Similarly, we found that generative AI was being used in an incredibly broad range of tasks. We asked people to select from a list of 10 which tasks they had used generative AI for, and then we asked them to rank those tasks in order of usefulness. The chart below shows what share of workers ranked each task in the top 2. Writing was ranked first (it was also ranked first for usage outside of work).

But the spread of the rankings was most striking to me. 8 of the 10 tasks got ranked in the top 2 by at least 10% of respondents. All 10 tasks were rated as useful (but not in the top 2) by at least 25% of respondents, which shows that people are using generative AI for lots of different things. This seems to support research and arguments by others that generative AI fits the definition of a “general purpose technology”, validating the clever paper title “GPTs are GPTs”.

How will generative AI affect labor productivity?

This is the trillion-dollar question. Here is an extremely speculative answer. In addition to asking about frequency of usage (weekly, daily, etc.), we also asked about intensity, that is whether people used generative AI less than 15 minutes per day, between 15 minutes and an hour, or more than an hour. It would be better to know the exact number, but people don’t keep track of these things, so we opted to ask for ranges.

We multiplied frequency times intensity to get some bounds on the total amount of work usage. Using the upper and lower bounds on the answers of workers who use it and assuming zero for those who don’t, we estimate that between 0.5 and 3.5 percent of work hours in the U.S. are currently assisted by generative AI.

We then looked at the productivity impacts estimated in five different experimental or quasi-experimental studies of generative AI access. The median estimate was 25 percent. Multiplying 25 percent times 0.5 to 3.5 percent tells us that generative AI could plausibly increase labor productivity (e.g. output produced per hour of human work) by between 0.125 and 0.875 percentage points at current levels of usage. For context, productivity has grown by about 1.5 percent per year since 2004. It grew at a 2.9 percent clip during the 1990s tech boom.

Don’t take those numbers to the bank! They assume that the published estimates from small studies would scale up nationwide, which you should not take for granted. It also assumes that usage “in the field” is concentrated in the same kinds of tasks that researchers decided to study, which seems broadly true, but it’s hard to know for sure. Think of these estimates as bounds on what is reasonable right now. They say nothing about the longer-run impact of generative AI on the nature of work, competition between firms, or larger impacts on the economy as the technology evolves.

If you made it this far, a story about how this survey almost didn’t happen

I hope you think our study is useful. It only happened by serendipity. I was giving a seminar at Vanderbilt in April, and I met with Adam. He told me how he and Alex had successfully designed the RPS, but that they might need to shut it down because they could not find any funding. I offered to fund it on the spot on the condition that we add some questions about generative AI, and we cooked up the survey in a matter of weeks. Our pilot was in the field by June, and our full survey in August.

Some of you may know that I recently wrote a column for The Atlantic provocatively titled “Break up Big Econ.” This example illustrates the exact point I was trying to make. Labor market surveys that establish important facts are incredibly valuable to the public, yet Alex and Adam could not find funding to continue their good work. Because I’m at Harvard I have access to more resources, including funding from organizations like the WalMart foundation (which partly funded the survey – thank you!) and others.

Public funding for surveys like this is nearly impossible to obtain. In fact, federal budget cuts are currently threatening the CPS itself. The BLS recently announced that it may have to cut the CPS sample by 5,000 households due to resource constraints. (I signed this open letter of support for BLS, and I urge you to sign it as well.)

We desperately need more public funding for the creation of new and important economic facts, and more generally to support economists who want to make useful practical contributions to the world.

As for us, we plan to field the RPS Generative AI supplement on a regular cadence so that we can understand how usage of this new technology is evolving. We will soon launch the next round, so post your ideas for what questions to ask in the comments!

We’ll be back to our usual schedule next week, when I’ll talk about machinists.

1. A recent study by Anders Humlum and Emilie Vestergaard found very high rates of ChatGPT usage among workers in eleven “highly exposed” occupations in Demark – these are jobs like software development and marketing where you would expect regular usage.

2. Keep in mind that my professional circle looks a lot like me – an aging, out-of-touch academic.

3. Full disclosure – from 2019 to 2022, I served on the Technical Advisory Committee of the BLS. It is an incredibly well-run agency and one of the crown jewels of U.S. data infrastructure, in my humble opinion.

4. We deliberately use the phrasing “outside of work”, because usage “at home” might be people working from home.

5. The CPS didn’t start asking about internet usage until 2001, so we also include data from the International Telecommunications Union (ITU), who collected global data on internet subscriptions in partnership with the World Bank. As you can see in the figure, the data line up well for the years that overlap.