GenAI in the Workplace

>> Steven Davis: Who uses ChatGPT, Gemini, Copilot, and other GenAI tools? How do these tools affect productivity in the workplace, and can they propel a lasting productivity boom? Our guest has been studying and thinking about these questions, and he has some answers. Welcome to Economics Applied. I'm Steven Davis, the Thomas W and Susan B Ford Senior Fellow at the Hoover Institution.

Joining me is David Deming. He's the Isabel and Scott Black professor of Political Economy at the Harvard Kennedy School. His research focuses on higher education skills, technology inequality, and the future of the labor market. You can find his newsletter, Forked Lightning on Substack. Welcome, David.

>> David Deming: Thanks for having me, Steve.

Looking forward to the conversation.

>> Steven Davis: As am I. So much of your recent research investigates the prevalence and impact of generative artificial intelligence in the workplace. So let's just start with a definition and some examples. What exactly is generative artificial intelligence, or GenAI for short?

>> David Deming: Yeah, it's an important question.

If you're going to do a survey or do some investigation of this, what do you actually define as it? In some sense, we're all using generative AI. If you've ever used autocomplete on your phone or your email or you search on Google for something and you get an AI answer, you're using the technology.

But we, in our work, we use a definition that's a little bit more restrictive than that. We decide and let me just actually read you the exact question from our survey, which we can discuss later. We say generative AI is a type of artificial intelligence that creates text, images, audio or video in response to prompts.

Some examples of generative AI include ChatGPT, Gemini, DeepSeek, MidJourney, whatever, and you can put whatever you want in there. And so we think about it as like you're using a program that is designed with, you know, large language model technology in the background.

>> Steven Davis: So, and much of what, much of what I'm going to ask you about and what, I presume much of what you're going to base your responses on is actually based on your own survey research.

>> David Deming: I should say too, Steve. I mean, part of when people talk about AI affecting the labor market, I think most people think about generative AI, which is in some sense just the latest incarnation of a trend that's been going on for longer. Some people would call the older generation predictive AI, which is basically just using big data techniques to make fancy predictions like, you know, which product are you going to buy next?

What price should I offer you? How do I optimize inventory management and all those kinds of things. So that technology's been around longer, but it turns out the same underlying thing you can apply to words. And that's why we have large language models that have taken the world by storm.

>> Steven Davis: Right. And there are other forms of AI that maybe aren't, don't have as much penetration in the workplace, but have potentially profound implications for productivity down the road. They're not our main focus today. So we're not about all AI, we're about generative AI and particularly as you just defined it.

But I did want to stress though, one of the reasons I asked you on this podcast is because there are lots of discussion, lots of people talking about AI, including gen AI, and many of the discussions are long on speculation and short on evidence. And what I like about your research is there's a real systematic effort to bring evidence to bear to the table so that at least when we're doing a more speculative discussion, that there's some evidentiary foundation for it.

>> David Deming: Yeah, that's the goal.

>> Steven Davis: That's the goal, right? All right, so let's get into some of it. So, you know, what do you learn from your surveys and other sources about the share of working age Americans who actually use gen AI tools in the workplace? I mean, how prevalent is it?

>> David Deming: Yeah, so before I tell you about the results, Steve, let me briefly explain our approach to it and why I think it's the right approach. So there's a lot of surveys out there that ask people, do you use generative AI? But they're often just based on what I would call convenience samples.

So you kind of send a survey out to a bunch of people on the Internet and you ask them, and the critical question is, are the people who are answering these questions representative of the average American worker or the average American citizen or citizen of any other country?

And, you know, it's hard to know. And so what we did in our survey was try to get at some underlying ground truth about this by creating a survey that is linked to the Current Population Survey, which is the survey that is the most important source of labor market information.

It's a federal survey administered by the Department of Labor where they send out, you know, questions to a bunch of people every month. And it started, if you remember the jobs day, the, you know, every month and the unemployment rate. It's the main source of labor market data.

And so we, you know, in, in the survey itself, for those of you who are like CPS nerds, you know, it has a quite baroque, complicated survey structure. So it asks a lot of Questions about, you know, are you unemployed? Why are you unemployed, or usual hours of work.

There's a lot, all kinds of things it asks. But so what we did was we contracted with a private survey company to exactly replicate the question wording, ordering, structure, timing of the cps. So we give our survey in the CPS reference week, we ask all the same questions in the same way, and then we tack on some other questions at the end about generative AI.

And what that allows us to do is say, well, and we, Sorry, we work with a survey company, we give them demographic quotas to ensure it's representative. But then we also, you know, so we can, we can see in the data if it is, but then we also say, well, if it isn't, we can weight the data to match the cps and then we can ask, okay, for the things we didn't give quotas for like hours worked, earnings.

Does it look like the people in our survey match the people in the cps? So we do a bunch of work in the paper to convince people that's true. And then we really want you to think about it as. Imagine that on jobs day every month, in addition to reporting the unemployment rate, we also reported the generative AI usage rate.

That's what the survey is trying to be. And you know, you never know, obviously, if you're getting the same people as the cps. But we're trying our hardest, Steve, to create something that we think of as truly representative. Okay, so that's my preamble.

>> Steven Davis: Right.

>> David Deming: Okay, so I appreciate all that and that's the right way to do it.

>> Steven Davis: It is worth noting the CPS itself, there are reasons for concern about representatives, but it's probably the best benchmark out there and you are benchmarking to it as best you can. That's how I think about your.

>> David Deming: That's exactly right. And you know, there's some larger question of whether you're ever going to get to the real truth.

But we're doing the best we can. And you know, I mean, I think it's comforting that having done all that, we get answers that are not that far off from a lot of the convenience samples. But that doesn't mean it wasn't worth doing, you know, so what do we find?

You know, we find that basically about 40% of people are saying they use generative AI at least, let's say once in the last month, like they are a user of generative AI. And of those, about a third overall say they use. They've used it at least once in the last week, either at work or outside of work.

We don't say at home because you could work from home. And then among those about a quarter, like 24% say they used it at work at least once in the last week, and about 1 in 10, 10% say they use it every day at work in the last week.

So pretty widespread usage. And there's a not insignificant share of people who are what I would call power users. They use it every day and they often use it for more than an hour on those days.

>> Steven Davis: That was kind of what I was going to ask about next.

I'm guessing that some of these people are just sorta dabbling.

>> David Deming: Yes, quite a few. And we try to get at that with our questions about intensity. So we say, first of all, we ask people, have you heard of generative AI? Okay, after we define it. And I think like 77% of people on our survey said they'd heard of it.

If they haven't heard of it, we don't ask them any other questions about it because we don't want to know if you're a user, if you haven't heard of it. Okay, so that's so like those people are just out. And among those people, we say, okay, if you've heard of it, like, do you use it?

How often do you use it? What do you use it for? And we get into a bunch of detailed questions that allow us to kind.

>> Steven Davis: When you set aside the dabblers, what What are people using it for in the workplace? Obviously, it's going to differ by occupation and so on.

But what are the important uses, as far as you can tell, of gen AI tools in the workplace right now?

>> David Deming: Yeah. So we ask, we give people a list of possible tasks, and it's kind of hard to come up with that because people do a lot of things at work.

But we had some that are things that a lot of people do, like writing. Like, you know, writing software code, interpreting and summarizing text or data. You know, all kinds of different things that we came up with. We used a bunch of tasks that are common tasks in a survey called the onet, which is a survey of what people do on the job, administered by the Department of Labor.

So we use tasks from O*NET, and the most common thing people use it for is writing. That's most common both at work and outside of work. So people even at home say they use it to write emails for them. And then also something that's closely related to writing, which is like interpreting, translating, summarizing text.

So, like, maybe I don't write any. Have it write an email for me, but I have it analyze an email for me. Or I say this is what I use it for a lot. Like, I'll write something, and if I'm writing a column and I got to get it down to 1200 words and it's 1300 words, I'll say, you know, cut this to 1200 words, and it's pretty good at that.

>> Steven Davis: Is that right? Okay, I got to keep that in mind.

>> David Deming: Yeah, no, it's actually great. Or if you say, like, you know, change the style of this, like, you can tell it, here's a column right in the style of Ernest Hemingway or something, and it'll be great at that.

Maybe you don't want to submit that, but the point is it's very good at taking something you already have and modifying it in some particular way. Almost better, I would say, than just writing from scratch. So I think a lot of people use it for that. And then, like, people use it a lot.

I mean the frequency of use for coding is lower cuz most people don't write software code. But people report very high time savings from coding, and a lot of our power users are people who use it to write software code.

>> Steven Davis: Just as an aside, it's not the main thrust of our conversation, but I've been working a lot on remote work in recent years with Nick Bloom, Jose Rio Barrero and others.

And one thing I. I hear from Managers in organizations that are remote work intensive and where there's a lot of asynchronous work activity that there's really a premium on clear written communication because you can't talk the way you and I are face to face. And if I say something you don't understand it, you make, well, a funny face.

A funny face. And then I realize I gotta elaborate. It's got to be clear out of the box and it can't be 10,000 words because nobody wants to read it. And what I hear you say now is actually there's a complementarity between these Gen AI tools and remote work, especially when it's in an asynchronous mode because it facilitates clear, succinct written communication.

>> David Deming: So I think people are using. That's exactly right. And actually there's an assistant professor, I think, at UCLA Anderson, named Gregor Schubert who's written a paper about this, about the complementarity between remote work and generative AI. AI, which

>> Steven Davis: I haven't seen that.

>> David Deming: Look that up. Yeah, it's interesting.

>> Steven Davis: Gregor Schubert.

>> David Deming: Schubert.

>> Steven Davis: Okay.

>> David Deming: S-C-H-U-B-E-R-T.

>> Steven Davis: T All right, I'll look that up. So, all right, so you already hinted at this question of usage intensity. So kind of where we're building towards is trying to assess some, you know, where we're going.

But I want to tell the audience so they understand some of the, where some of the questions are coming from. We're trying to assess the overall productivity effect along ultimately on the US Economy now and going forward. And to do that, we need to know not just who's using it, but how intensively they're using it and put that together with some information about the productivity impact of when they're using it.

>> David Deming: Yeah.

>> Steven Davis: So what can you tell us about from your survey evidence about the intensity of usage of Gen AI tools in the workplace today?

>> David Deming: Yeah, so what we do. You're exactly right as to where we're going. I mean, I think that's the question we had in mind when we designed the survey.

And you know, so the intensity of usage varies a lot by person and by occupation. Not surprisingly, you know, as I mentioned, some people, you know, 1 in 10 roughly use it every day at work. And then another decent share, you know, use it at least one day or more than one day, but not every day.

Then some people are very occasional users. So we have that. That's kind of like how many days you think about that as like filling out how many days of the week or the work week do you use it and then there's the other bar, which is okay, on the days you use it, how much you use it.

>> Steven Davis: Right.

>> David Deming: And so we ask people that too. We say, on the days you use it, do you use it? You know, one to 15 minutes a day, 15 minutes to an hour, more than an hour, you know, an hour to three hours, three hours or more.

We have different categories. And so then that's, if you take those two numbers, how many days do you use it? And then when you use it, how often do you use it? You can kind of add those up, multiply them together actually, and get a sense of like how many minutes of every work week.

Let's say, are you using AI?

>> Steven Davis: Yeah.

>> David Deming: And then, you know, of course a lot of people are saying zero because they don't use it at all or they've never heard of it. But some people are using it all the time. And so you can add, basically add all that up and say, okay, if all the work hours currently, you know, being, I don't know, worked, I guess in the US labor market right now, what share are being assisted or what share are being, you know, co produced with generative AI?

And we asked our questions, we have a range. We don't say is it 37 minutes or 38 minutes, we say between 15 and 60. So we basically use those as lower bounds and upper bounds. So like, if you said 15, suppose it was actually 15. If you said between 15 and 60, suppose it was actually 60.

And we estimate that between 1 and 5% of all work hours in the US right now are being assisted by generative AI. And that includes the.

>> Steven Davis: How does that vary across industries and occupations? To the extent you can say.

>> David Deming: It varies hugely, not surprisingly, I'd say the biggest industry is like information services, you know, and this is not that surprising.

You know, big companies that are technology intensive, you have usage rates that are, you know, two and a half to three times higher than that. And then in some types of firms, like personal services, occupations and leisure and accommodation, kind of like retail sales or working in hotels, people are really not using it very much at all.

They are using it some, but they're not using it very much at all. So it varies a lot. I will say the one thing that I was a little bit, if you look at my occupation, the highest usage rates are in management, you know, business, stem, whatever. That's not surprising.

But actually, even in like blue collar work, it's like 20% of blue collar occupations say they use AI at least once in a While, and, you know, something like 1 in 6 or 1 in 7 are using it once a week. So for me, I was quite surprised by the breadth of usage across occupations.

And in the paper, we actually compare that systematically to personal computer usage because the CPS, you know, 20 years ago, started asking questions about PC adoption. So you can look and you can say, well, actually in like software occupations, everybody was using PCs right away, but in half the occupations in the US labor market, basically nobody was using a PC until much later.

And relative to that, generative AI is much broader.

>> Steven Davis: Okay, and I don't know whether you ask about this in the survey, but do people have a sense of whether when they use these GenAI tools, they're mostly just speeding themselves up or they're actually improving the quality of what they do?

>> David Deming: Yeah, so we asked a lot of questions about that. We tried different things in different waves of survey. But the thing we settled on, which I think is the best we've done so far, is we ask people questions about how much time they save and we say specifically, you know, in these tasks, in, in the times you used it, how much more time would it have taken you to do those tasks if you had not had generative AI?

And we use that to back out the time savings from the technology. And so we ask people, you know, and there's a pretty wide distribution. Some people say it saves them four or more hours a week. Some people say it saves them, you know, less than an hour.

Some people say one to four. It's actually pretty variable. But we can then take those time savings and multiply them times the first numbers I said. So it's like what percent of hours are assisted by AI and then how many hours are saved by that? And what we estimate is that for users of AI, they're saving on net, about 5.4% of all work hours using AI, they estimate.

And then if you include non users, about 1.4% of all work hours are being saved by generative AI, which, you know, it. It's actually kind of a big number, I think, but depending on your prior belief.

>> Steven Davis: Yeah, did you ask them what they do with the time savings?

>> David Deming: Yeah, so that's a great question. That's what we're doing in the next wave. So I should say we're doing this every three months and we're always at That's a key question. I think the reason is that, you know, we'll talk about productivity, but I think a really important piece of this that's missing from the conversation in my view is that because worker adoption is so far ahead of firm adoption.

So like 40% of people are saying they're using it, but if you ask firms, it's like 6, 7% have a formal AI use policy. So what that's telling you is a lot of people are using it. Not, they're not being prohibited by their firms, but it's not part, they're kind of using it on the sly.

Okay. So your boss says to you, I need you to write the memo about topic X, have it on my desk in two days. And then you use AI, you write in two hours and you go to the bar.

>> Steven Davis: Exactly.

>> David Deming: Well that's, that's great for you, but it's not going to show up in the productivity statistics.

>> Steven Davis: Right, but it's still, it's still a value. It's still a value.

>> David Deming: Yeah, I know it's still value. But like whether, I guess that's what I mean, whether it shows up in the productivity data depends on whether firms internalize their, those expectations of what you can now do with technology.

And they say, okay, now I want it in two hours instead of two days. Then, then you'll see productivity gains.

>> Steven Davis: That's interesting. So, well it, there, there's. It also, it's complex the way it might play out in terms of compensation.

>> David Deming: Absolutely.

>> Steven Davis: Both cunary compensation and non-pecuniary forms of compensation.

You just don't have to work as hard. And I remember when I was stocking shelves, working in the grocery stores, I worked extremely hard in the first few weeks in the job. And every time I figure out how to do something more efficiently, part of it went to me working less hard, part of it went to higher productivity for the employer.

>> David Deming: And that's kind of fundamental to every now technology. Right. Sorry, Steve. Sorry to interrupt you. It's kind of fundamental to every technology and I think different technologies have fallen on different ends of that spectrum. So like people talk about the productivity impacts of mobile phones like it wouldn't surprise me if it was negative.

Right. The welfare gains are very huge because now you can watch videos on your phone and those are consumer facing technologies. They're not really intended to increase productivity.

>> Steven Davis: All right, so let's, let's go back to the main path.

>> David Deming: Yeah, sure. Sorry.

>> Steven Davis: So no, this was fun.

It was great. It's good to think about these things, but I want to get back to the productivity. Yeah, you know, the big picture productivity story. So, you've told us how you measure usage intensity. You have some way to connect it to time, the time savings aspect of productivity.

It's, there's still a question of how they use those time savings. But how did you go from your data on usage intensity and time savings to coming up with some broader assessment of productivity effects, both particular jobs, people, industries, but then the economy as a whole. So, we're trying to, we're trying to get to the billion-dollar question of is this really going to propel a productivity boom that matters at the macro level?

>> David Deming: Yeah, yeah, I think, bottom line, that's the important question. I want to say before I tell you what we did or speculative calculations were, I place a very wide confidence interval on this. I think anybody who says they know the answer is very spec. We're just trying to not even say what the answer is, but just give some boundaries on what it could be given what we know.

So that's what the exercise. Highly speculative. So here's what we do. And actually we do something that's analogous to a widely discussed paper by Daron Asimoglu called the Simple Macroeconomics of AI, where he does something similar, which is you just say, okay, well for any new technology, let's say generative A, like how often are you using it, in what tasks are you using it?

And how big is the productivity gain from using it? Okay, so it's kind of like, it's called Holton's Theorem. It's basically like you take how intensively is this new technology used? In what task is it being used, how valuable are those tasks? And then how much of a gain does it give you?

So it's almost, it's an accounting exercise really. And so, what he does in his paper is he says, there's a paper by Elundu et al. Where they use a bunch of expert assessments of, you know, which tasks are potentially replaceable by AI. And then he says from that paper, the estimate's about a 20% occupation exposure share.

And then he takes another paper that says, well, of those occupations that are at risk because that's an occupation level comparison, what share of those tasks in those jobs? Let's say, you know, coding if you're a software developer, versus like having meetings which can't be automated. You know, what share of the task can be automated by AI?

That estimate is 23%. So you multiply 20% times 23%, you get 4.6%, then you take that, and you multiply it times the productivity gains from AI. And so he reviews six different studies that show that when people use AI, they get about 27% more productive. So you multiply that 4.6% times the 27%.

And then you say, well, if you want to look at total factor productivity, there's labor and capital. The labor share is 0.57. Okay, so that's a lot of math, but you're basically taking 4.6% times 27% times 57%. And the answer is, you get a plausible impact on TFP of about.07 percentage points over 10 years.

So that's.

>> Steven Davis: It sounds complicated because we're just, you don't have the blackboard here or the whiteboard, but the basic arithmetic is fairly simple.

>> David Deming: It is.

>> Steven Davis: It's just multiplying and adding up a bunch of stuff where you take account that certain industries and sectors are account for a bigger share of economic activity than others.

And you're doing the sectoral level calculation and then kind of adding it all up.

>> David Deming: That's exactly right. And like, if you just want to think about it not with precise numbers, you're saying about a fifth, you know, about a fifth of occupations weighted or exposed to AI.

In those occupations, about a fifth of the tasks you do are exposed to AI. You multiply those two together and then you say, well, you get about 25, about a quarter more productive. You multiply all those together and you get that number. So it's basically just multiplying a bunch of these numbers together.

Now, what we do in our paper is we arrive at an estimate that does the same thing. It's the same kind of multiplication exercise, but it's all using our survey data rather than using expert estimates. So, we say, what share of tasks are AI? You know, what share of the time at work are you using AI?

We know that from our survey. So we multiply all that out, ??

>> Steven Davis: right?

>> David Deming: Then we take the time savings estimate and we say, okay, in the task, you're using it for how much time are you saving? And we use that to generate our own productivity calculation rather than using other studies.

And we get a number that's a little bit higher. It's 33%. So, it's 27%. And so. And then our. Our time usage numbers are actually very similar to the exposure measures that Acemoglu has. And so we get an answer that's like 1.1%. Basically, it's the bottom line. So, Acemoglu gets 0.7, we get 1.1.

>> Steven Davis: This is your. Just look at this 1.1% number. This is. I want to make sure I understand it.

>> David Deming: Yeah.

>> Steven Davis: This is your estimate, ballpark estimate of how much GenAI has contributed to the aggregate labor productivity.

>> David Deming: Yes.

>> Steven Davis: Since say 2023, when these technologies really came on.What's the time period we're looking at here?

>> David Deming: So, I would say slight slide modifications. I would say it's our estimate of if the labor market were perfectly competitive. Meaning firms and workers perfectly internalized things and everybody was producing more as a result of having AI. How much of an effect could it have at current levels of usage?

It's not really so much about like the past. It's saying right now, okay, but how much more productive would it be making people?

>> Steven Davis: All right, but the reason I picked 2023 is because these tools weren't really in use much. That's fine as you define them. So I'm thinking about over the past two years.

I take your 1.1%, 1, 1.1% number. We've got a 1.1 percentage point extra boost in productivity growth from this source alone. And that's non trivial.

>> David Deming: I personally would say it's an upper bound more than a lower bound. Because you know, again, this thing I mentioned earlier, like if you get much more productive, let's say you get a third more productive.

But, but your employer doesn't even know you're capable of that. There's no guarantee they're then going to ask you for a third more output. If you're salaried. You're taking some of the benefit.

>> Steven Davis: Yeah. And here it depends on exactly how we want to think about productivity.

>> David Deming: Exactly.

>> Steven Davis: So, but, but as it would show up, I think what you're. Where you're coming from. As it would show up in the traditional productivity statistics.

>> David Deming: Exactly.

>> Steven Davis: The government authorities produce. It's probably an upper bound for the reason you just said that workers. Workers are taking some of the benefits in terms of their own utility rather than more output per unit time.

>> David Deming: Yeah, I think right now. And that's not state, that's not a stable equilibrium, right? So like eventually firms are gonna rise up and start demanding more and so then.

>> Steven Davis: Yeah, well, I've made very parallel arguments with respect to the unexpected surge in remote work that initially a lot of the benefits went to workers, especially people like you and me, who could work remotely effectively.

And yet, there was no change in our compensation immediately. So it's kind of like all of a sudden we got this big benefit without that went mostly to the worker and there were adjustments over time.

>> David Deming: I think it's exactly analogous.

>> Steven Davis: Yeah, I see the logic of your relatively Icemoglu et al approach.

And you've got more data, it's more granular, you can do a more ground up approach. Then it's not a criticism of what they did, it's just they didn't have the same data foundation. But there's a different kind of question which you can ask, which might be to say, okay.

Let's imagine a few years down the road where everybody who can benefit from AI at the individual level and the organizational level has figured that out. Because innovations tend to diffuse slowly over time. We know that for many studies in economics. So if I wanted to kind of get where we might go over the next 5 to 10 years just with the gen AI tools.

I might then want to go back to the expert approach and say, okay, when this has really been widely adopted, how much can we get in terms of productivity growth? Can you speak to that?

>> David Deming: Yeah, I think that's a very nice point, Steve. And I think just underscore what this calculation is, which is it's a very point in time, partial equilibrium, to use an economist term, calculation.

It's not thinking about, well, now that this basically intelligence is just more or less freely available, how do I redesign production in my firm? How do I hire differently, how do I learn differently? It's not incorporating any of that, it's just saying at current levels of factor utilization in the economy, what is this capable of?

And we know from many the history of technological change that's not even the first order impact of these technologies. It's sort of like saying, if we did a calculation like this back in 1987 with a personal computer. We would be missing a lot of the ways that our lives are fundamentally different because everybody has a computer available to them and so that's gonna be true, I think with AI.

Just to think concretely about it. A lot of people are talking about, startups are much easier now because you've basically got like, I think the best way to think about the AI is like it's staff. You've got an AI that can do a lot of things for you at relatively low cost and so you don't need to staff up as much when you're a startup, which means you can be leaner for longer.

And you have a longer runway in which to succeed. And so all those things are not in the calculation, but could have really, seriously big impacts on productivity down the road.

>> Steven Davis: Okay, so you, you're not gonna venture a number, it sounds like.

>> David Deming: I, no way I think.

>> Steven Davis: I was actually asking, you a little less ambitious question which is more closely tied to what you're currently doing. So you're, you're talking about all of the ways we will redesign the workplace and how that will, and then facilitate new business entry and so on. So let's take that off the table but just take your current penetration rates are and say, okay, well what could we expect 10 years from now in the current, in the current kind of structure of the way we do work?

If we get three times as much usage intensity as you see currently, well, then we get in here, we're still in this kind of approximate Halton's theorem environment, but that's fine. We get three times as much benefit as I understand.

>> David Deming: Well, if there are no diminishing returns, I mean like maybe the technology is now being used in its highest value ways.

Maybe the software, I don't know, it could be three times as much but there are reasons why.

>> Steven Davis: You've anticipated I was going to come to reasons why it could be more or less than the simple calculation might suggest. And you've just pointed to one where might be less.

And that's because this goes back to a famous paper by Zvi Grillich's and with hybrid corn and the rollout of hybrid corn technologies. He shows that it diffused across the United States to the places first where the profitability of adopting it was highest. In this case, we would expect the take up of, of Gen AI to be quickest, where its productivity gains were the highest.

That's your point. And so that's one reason we might, if those if that kicks in very quickly, then we might not get that much more than we've already got.

>> David Deming: You can tell the story the other way too.

>> Steven Davis: Well, you sketched one earlier, but go ahead, tell the story the other way now.

>> David Deming: I think the other story would be like somebody said, I forget who said this, but it's a nice line. There are very few AI sized holes in the economy. So it's like right now it's just plug and play, like you're just sticking it in there. But actually if the AI gets a little bit better and you can automate a higher share of the tasks, there may be complementarities with whether it's worth it to redesign your organization.

So it's like, I mean, one analogy would be if you go way back to the mechanization of farming, okay. People, when steam power came online, people had, people invented steam tractors. And so the question was, should I replace my horses and mules with the steam tractor? It's like, on the one hand, it relieves the power constraint to feed these things as much.

It's much lower cost to get the same power. But the steam tractor weighs like 2,000 pounds. It breaks down all the time. It doesn't burn very well, right? And so you're well, it replaces half of the things I need at good cost, but I still need the horses and mules for these other things.

So it's not really transforming my operation. But then when you invent like the general purpose, smaller, more mobile tractor, then I'm willing to give up and then I'm really willing to change. So there's a sense in which once the AI reaches a reliability or a productivity level where it replaces enough that you're willing to make serious changes, it could be bigger.

So that's a story.

>> Steven Davis: And that, that's true at the individual level and the organization level.

>> David Deming: Yeah.

>> Steven Davis: There's another, there's another force that goes in the same direction, which is as we use these tools at the individual level and the organization level, we will probably get better at using them.

>> David Deming: Yes.

>> Steven Davis: And recognizing what they're most useful for.

>> David Deming: I think that's right.

>> Steven Davis: Yep, so that also cuts the other way. So I think we're, we're basically as good economists sketching out. Well, on the one hand it could be more. On the other hand it could be less than you might say.

And explaining why there's good grounds for uncertainty about exactly how this will play out.

>> David Deming: I think that's right. The only thing I'd say I'd add to it just in our defense of the one armed economist legendarily is that's the reason to collect data like this is that actually some of the things we've said on either side can be directly tested in data.

So for example, if the adoption curve increases over time, we want to know that right away. That tells us that it's being used more. It could flatline, we could have more power users in some industries rather than others. And so that's why it's so important to get on the ground floor.

This is some basic data as soon as possible because I think the trends are gonna tell us in some sense even more than the levels about where this is going.

>> Steven Davis: I think there's another broad lesson here, which is this is an example of where markets are extremely powerful relative to some top down planning or design approach.

Because there's a million and one different permutations of ways to use these technologies. Many of them will be tried and many of them will not perform very well in a market setting. What doesn't work, people move away from quickly. They keep trying to figure out what works, what doesn't work, the profit motive, or at least some notion of a business success of what drives them.

So the reason I stress that is I sometimes get worried about what strike me as. Heavy handed, potentially heavy handed regulatory approaches that want to slow all this down because we don't know what will happen and something risk something bad might happen. I think that's the wrong approach in this, in this setting.

>> David Deming: I mean the bad thing that can happen is we might slow it down like that could.

>> Steven Davis: Exactly, exactly.

>> David Deming: So yeah, I totally agree, I totally agree with that.

>> Steven Davis: We don't know exactly how this is going to play out but all the more reason why we want, you know, let's let 1001 different approaches bloom and it's back to, you know, something you said earlier about how these tools facilitate.

I think they facilitate small scale business entry and the same is true to some extent by the way of remote work because you can now start a small company in a small town and you don't need all of the specialized labor inputs right around your location. You can get some of them, you can get some of these services remotely.

So again it's interesting to me there's some parallels between, as I think about it, between these gen AI tools and the shift in some activities towards remote work in that they both facilitate much broader but small scale business experimentation and entry. And that's likely to be a great thing for the economy over the long term because some of these micro experiments will turn into the huge business success stories of the next generation.

>> David Deming: I think that's absolutely right. It's a great point and I think it's actually here's another story for why it might accelerate. I could see a lot of this happening faster next time the economy enters a recession because right now we're sort of at full employment. Companies are really reluctant to make big changes.

But as soon as something happens where people really need to adjust because of the competitive pressure, that's when you'll see adoption speed up if history is any guide. So yeah, this could, and that's another reason to keep things as open ended as possible so that when we hit the next negative shock in the economy, we can be flexible enough.

There's not some rule telling us we can't use AI for certain things to restart the economy with small businesses or whatever.

>> Steven Davis: Right, right. So maybe that. So are there any other, anything else on the horizon that you see that might actually inhibit this process or slow it down or prevent us from realizing these kind of gains that we see out there?

We don't know exactly what form they'll take. But I think it seems like you and I are both on the same page. There's a lot of potential here.

>> David Deming: I do have A second order of worry, which is maybe a bit outside the scope but worth talking about, which is, you know, I think it's.

I know there's been a bunch of studies of AI that show that it kind of like brings up people who are novices. So like it could be inequality reducing. I don't really believe that. I think that in those studies what you see is AI helping you with some specific task where the task is well defined.

But the studies I've seen, like this paper by Aidan Tony Rogers about the impact of, he studies the impact of generative AI usage in scientific discovery. And then there's another paper where they give AI help to entrepreneurs. And in those cases what you see is the people who are more capable at baseline know how to use the technology better because it's kind of like using it in the wild.

Here's a technology, you can use it for whatever you want. And the thing I worry about is that AI is such a crutch in education, like a lot of our students are using it. And I worry that there's some problem where today it's really helpful to use AI to get your assignments done or whatever.

But actually the people who can use it best are the ones who have deep domain knowledge because they know when the AI is right, when it's wrong, what to use it for. They're better at experimenting. And so I do worry a little bit that in the education system human capital formation will suffer because of AI, because we haven't really worked out.

The education system is kind of sclerotic and we haven't really worked out like how to adapt it to AI. So that's a second order work.

>> Steven Davis: I hear your concerns. I guess I have a slightly different take and I'd make a distinction between domains in which AI is really a complement to innovative activity, where I share your view that it may just may create an even more, a greater opportunity expansion for people who are most skilled.

>> David Deming: Yeah.

>> Steven Davis: But then there are, if you move away from the innovation part of the economy and more towards where it's really routine, it's execution of routine tasks efficiently. Those are the areas where I think there's been this raising up the bottom. So think about things like customer service, especially remote customer service.

If you are providing even simple technical help for some appliance for a customer or their electronic system or their computer remotely, I think it really does help if you could bring up the, the least capable customer service employees to the median, which is something that AI seems, AI tools seem to be useful to do.

So I guess I'm not quite as pessimistic as you that I see. I do think it's gonna be heterogeneous that in some activities, especially where it's more about just routine execution, performance rather than innovation, that AI could actually be a leveler rather than something that drives more inequality and in productivity and output.

>> David Deming: Well, I hope you're right, Steve. I mean, I guess the critic, the thing I would say, and if there were like some, you know, tech guys here, they would say, well, those jobs are just going to be replaced, you know, And I, I'm. I'm not somebody who thinks AI is gonna take all of our jobs, but if there's one job it is gonna take, it's probably like back office, customer service, finance, billing type stuff.

So, I guess the counter.

>> Steven Davis: We'll see, we'll see soon.

>> David Deming: Exactly.

>> Steven Davis: See, it depends on how much people dislike thinking they're dealing with AI.

>> David Deming: Yeah, true. And how good it can.

>> Steven Davis: If they can figure out some way that I don't, that I can call my health insurance company and I don't have to spend 30 minutes on the phone talking to five different people.

I'm all for AI, but I don't see it yet. I don't see it yet.

>> David Deming: Yeah.

>> Steven Davis: So let's wrap. This has been a fun discussion, but we can't wrap up with asking the really important question, which is why is your substack newsletter named Forked Lightning?

>> David Deming: Great question.

I'm glad you asked. So the newsletter is really my chance to write about, you know, the research that I'm doing and other people are doing in my own voice and to kind of translate it for an interested lay audience. That's really what I'm doing in it. And I called it Forked Lightning because Dylan Thomas is my favorite poet, and he wrote a poem called, Do Not Go Gentle into That Good Night.

And in the poem, he's talking about, you know, the regrets people have at the end of their lives. It's a poem that he writes when his father is dying. And, you know, he says, wise men, you know, though they know dark is right because their words have forked no lightning.

They do not go gentle into that good night. And what he's saying is, people who have a lot of knowledge but don't use it to create more impact and good in the world have regrets at the end of their lives. And so that's why I call it that. It's like, let's fork some lightning. Let's actually make an impact.

>> Steven Davis: I see.

>> David Deming: Let's fork my light peculiar knowledge. Let's try to put it to some use, so.

>> Steven Davis: Okay, I like that. Well, it's a famous poem, and even. Even uncultured people like me were aware of it, but I.

I had to go and look at the fourth lightning passage.

>> David Deming: Yeah, Yeah.

>> Steven Davis: I didn't remember that passage. It's the. Go. Don't. Don't, you know, rage. Rage against the.

>> David Deming: Yeah, that's the thing.

>> Steven Davis: That's the passages that were already in the-

>> David Deming: And I love the idea of it, because we're tenure professionals.

>> Steven Davis: No, it's great.

>> David Deming: Need a rational laurels. And I'm gonna rage a little bit, that's the idea, we'll see.

>> Steven Davis: Yeah. Yeah. All right. All right. I like that. Okay, so I'm gonna give you the sign off. Live long and fork lightning.

>> David Deming: There we go, I love it.

>> Steven Davis: That's the new sign off. Live long and fork lightning.

>> David Deming: Live long and fork some lightning, all right? That's great. Thanks so much. This is fun.

>> Steven Davis: Thanks, David. Take care.