Before conducting a recent survey of Americans to figure out how much they're using generative AI, Harvard University economist David Deming says he was solidly in the "AI skeptic" camp. That is, he was skeptical that the explosion of generative AI would offer sizable benefits for the U.S. economy anytime soon. Now, however, he says he's more optimistic.
"I was very surprised at the numbers in our survey," Deming says. "And it sort of made me think that AI is gonna be a bigger deal than I would've thought."
The study, Deming says, was motivated by questions over whether and how much Americans are using generative artificial intelligence. Doing what economists tend to do, he and his colleagues, Alexander Bick and Adam Blandin, wanted to get some good data.
They modeled their survey after the Current Population Survey (CPS), which is sponsored by the U.S. Census Bureau and the U.S. Bureau of Labor Statistics. Conducted monthly, the CPS is sort of the gold standard for surveys. It's how we know things like the unemployment rate and the state of the labor market.
In short, the survey that Deming, Bick and Blandin conducted is high-quality. It has a large sample size. It's nationally representative. And they even included some of the same questions that the CPS asks, so they could cross-reference their survey with the CPS and make sure their numbers were good. They conducted their survey twice, in June and August 2024.
Deming says he was shocked by the results. He and his colleagues found that almost 40% of Americans, ages 18 to 64, have used generative AI. And a sizable percentage seems to use it regularly. In their August survey, the economists found that more than 24% of American workers had used it "at least once in the week prior to being surveyed, and nearly one in nine used it every workday."
Even more surprising, Deming says, usage of AI seems to be pretty much across the board. He expected that younger and more-educated Americans would be the biggest users. They confirmed that's the case. "But we even found that 22% of blue-collar workers say they use AI, and usage rates were above 20% in every major occupation category except personal services, where it was like 15%," he says.
This rapid rate of adoption, the authors point out, is much faster than the speed at which Americans adopted personal computers and the internet. That may not be very surprising though, considering that personal computers cost thousands of dollars in the 1980s and 1990s and you had to get physical hardware, hook it up, figure out how to use it, etc. Even with the internet, people had to buy modems, get subscriptions, figure out how to "surf the web" and so on. Generative AI is more plug and play, often free or has a low-monthly subscription cost, and has a user interface that is familiar to anyone who has ever used Google.
What does this mean for the economy?
In a recent two-part series on The Indicator, Darian Woods and I debated whether AI is overrated. Given the tremendous uncertainty about the impact that AI will have on the economy — as well as our reluctance to make predictions that might get us laughed at in a few years — we decided to obscure our personal feelings on the matter. We flipped an AI-generated coin. I got "AI is overrated" and wrote an episode and a newsletter making that argumentative case.
In the course of my research, I discovered a study conducted by the Census Bureau this year that found that only about 5% of American businesses said they had used AI in the previous couple of weeks. I used this as evidence that use of generative AI was pretty pathetic, especially considering all the hype around it.
How does Deming square his results with that finding?
For one, Deming says, he and his colleagues found that actually Americans report they're using AI more during their personal time at home than at work.
Hmmm. OK. Putting my “AI is overrated” hat back on, this makes me think that much of AI use is for pleasure rather than for productive work purposes — which would suggest that AI's impact on the economy will be limited.
The sort of economic dream for AI would be that it would rescue the U.S. economy from its long period of slow productivity growth. Productivity growth — meaning workers being able to produce more in less time — is the magic sauce of rising living standards. And recent technologies have been pretty disappointing on that front.
I mean, look at the smartphone. If I told you back in 2006, the year before the iPhone was released, that we'd soon all have supercomputers in our pockets, able to search the internet, give us precise directions to anywhere, send emails to and do video calls with co-workers and clients, order basically any product or service, translate languages and on and on — you might think we'd see an explosion in productivity. But smartphones seem to be more a tool for pleasure and distraction than an incredibly impactful work tool. We haven't seen a huge boost in productivity growth since its mass adoption well over a decade ago.
Deming, however, stresses that people are using generative AI for work. He says their survey shows “about 1 in 4 people used it at least once in the last week at work" (versus "about 1 in 3" who used it in their off time). When it comes to how people use AI to help their work, respondents indicated that writing, interpreting and administrative support were the most helpful (although a significant share also said they were using it for other tasks, including coding and support when dealing with customers and co-workers).
Squaring his study with the one from the Census Bureau, Deming says he thinks there's a "big gap between a company's formal policy on AI use and workers. I think quite a lot of the use is under the radar, like you've gotta write an email to your boss and you just use it to write the email more quickly and you just use ChatGPT to do it. But you don't tell anyone." Deming says many of his students use generative AI this way.
We reached out to Massachusetts Institute of Technology economist Daron Acemoglu — whom, these days, you might call the AI skeptic in chief — about Deming and his colleagues' study. Does the fact that they find that a large percentage of Americans are using generative AI change his perspective at all?
"My concern with their numbers is that it does not distinguish fundamentally productive uses of generative AI from occasional/frivolous uses," Acemoglu said in an email. "If you are curious about what ChatGPT would say to introduce one of your guests and use it to get a sense of that, you are a user, but this is not fundamentally integrated into the NPR production process. Worse, if you asked me whether I'm using generative AI, and if I'm being honest, I will have to say yes, because when I do searches, now I see generative AI-generated output. But this has zero impact on me. The productivity improvements would require fundamental integration and reorganization of production processes, together with complementary investments in organizational capacity and worker skills — not gimmicky, frivolous uses. We know ChatGPT has about 200 million unique monthly users, but the question is how many of them are using it in a way that will lead to significant productivity improvements/cost reductions. I don't really know the answer to that question, and I don't think we find the answer to that question from this paper."
How much will generative AI increase productivity?
To try to predict how much AI will boost productivity growth, Deming and his colleagues did some rough calculations. They looked at five randomized studies that analyzed how much use of generative AI increases productivity in various work settings. They picked the productivity-boost number in the middle of those studies, which is about 25%. They then multiplied that, Deming says, by "our estimate of the number of work hours that are currently being assisted by generative AI" (which, they calculate, is somewhere between 0.5% and 3.5% of all work hours in the U.S.).
They ended up estimating that generative AI will result in an "increase in labor productivity of between 0.125 and 0.875 percentage points at current levels of usage." That may not sound like very much. But, Deming says, consider that productivity growth over the last couple of decades has "been about one and a half percent per year. So if you take that 1.5% and you add this, that's actually a pretty big increase."
Interestingly, Acemoglu has also done some rough calculations of the potential impact of generative AI on the economy in the near future. And despite using a different method to calculate the potential productivity boost from AI, his estimates are actually pretty similar to this recent work from Deming and colleagues.
Yet, Acemoglu and Deming put different spins on the results. Acemoglu seems to be reacting more to all the media and industry hype around generative AI — and, he's like, we're not witnessing an economic revolution by any means. Deming looks at the potential productivity increase and emphasizes more that it looks like generative AI could make a meaningful impact on the U.S. economy.
"Is it gonna lead to 7% productivity growth? No, probably not," Deming says. "Not the way it's currently being used, but it could add a little bit, and every little bit counts. Like that's millions and millions of dollars of extra GDP growth and rising living standards. That really counts."
Deming and his colleagues have plans to continue conducting their survey in the future. It's worth noting that they conducted their last survey before Apple released an iPhone that has ChatGPT integrated into it.
Pretty soon, he says, “because AI is gonna be embedded in so many things, like the iPhone, it's gonna be hard to even ask people, "Do you use AI?"
Copyright 2024 NPR
