Scaling Laws: How AI Is, Will, and May Alter the Nature of Work and Economic Growth with Anton Korinek, Nathan Goldschlag, and Bharat Chandar

[Intro]

Alan Rozenshtein: It is the Lawfare Podcast. I'm Alan Rozenshtein, associate professor of law at the University of Minnesota and a senior editor and research director at Lawfare.

Today we're bringing you something a little different: an episode from our new podcast series, Scaling Laws. It's a creation of Lawfare and the University of Texas School of Law, where we're tackling the most important AI and policy questions, from new legislation on Capitol Hill to the latest breakthroughs that are happening in the labs. We cut through the hype to get you up to speed on the rules, standards, and ideas shaping the future of this pivotal technology.

If you enjoy this episode, you can find and subscribe to Scaling Laws wherever you get your podcasts and follow us on X and Bluesky. Thanks for listening.

When the AI overlords takeover, what are you most excited about?

Kevin Frazier: It's not crazy. It's just smart.

Alan Rozenshtein: And just this year, in the first six months, there have been something like a thousand laws.

Kevin Frazier: Who's actually building the scaffolding around how it's going to work, how everyday folks are going to use it? AI only works if society lets it work.

There are so many questions that have to be figured out and nobody came to my bonus class. Let's enforce the rules of the road.

[Main episode]

Kevin Frazier: Welcome back to Scaling Laws, the podcast brought to you by Lawfare and the University of Texas School of Law that explores the intersection of AI, policy, and, of course, the law. I'm Kevin Frazier, the AI Innovation and Law Fellow at Texas Law and a senior editor at Lawfare, joined by a trio of economists.

First, there's Anton Korinek, a professor of economics at the University of Virginia and newly appointed economist to Anthropic’s Economic Advisory Council. Second, there's Nathan Goldschlag, director of research at the Economic Innovation Group. And third and finally, there's Bharat Chandar, economist at Stanford Digital Economy Lab.

Today we're tackling the question that's dominating headlines and earnings calls: how is AI disrupting the workforce?

As companies like Amazon announced staggering layoffs, it's unsurprising that more and more Americans are wondering when they may find themselves on the wrong end of a CEO's effort to adjust the company to the age of AI.

This trio of experts is well suited to take on this important topic, and with that, giddy up for a great show.

So everyone knows when they pull up the latest edition of the Wall Street Journal, or the New York Times, or really basically any newspaper these days, there's seemingly some headline about how AI is causing some degree of economic turmoil, whether it's job displacement, whether it's the quote unquote AI bubble, or whether it's just the future of the economy itself. And that's why I'm so jazzed to have three incredible economists with us today on Scaling Laws.

Anton, Bharat, and Nathan, thank you so much for joining.

Anton Korinek: Great to be here with you.

Bharat Chandar: Thank you.

Nathan Goldschlag: Happy to be here.

Kevin Frazier: Bharat, let's start with you. In terms of economic consensus, what is the sort of status quo among economists, which is a lot like saying, you know, tell me about how frequently lawyers agree on the interpretation of the Constitution, but for our audience, can you kind of outline what points of consensus and consensus exists among economists?

Are we seeing a sort of doomer camp of economists, accelerationist economists? What does it look like if we were to draw a map of how economists are thinking about AI?

Bharat Chandar: That's a great question. There's a couple angles to that. There's the productivity angle, I think, and then there's the labor angle.

So I guess I'll start with the productivity angle, because I think that informs how we think about some of the labor market impacts. On the productivity side, I would say that the median economist is probably more skeptical of productivity impacts than the median technologist as you might expect.

Now, that said, there are definitely economists kind of across the spectrum and, you know, wonderful people like Anton are kind of modeling the potential impacts of transformative AI and how that might reshape the economy, both on the productivity side, on the safety side, on the labor side, et cetera.

On the labor side of that equation, I think we can both talk about future impacts and current impacts. My view of what we're seeing in terms of current impacts is that overall––and this is suggested by a few studies including by Nathan, myself, this team at the Yale Budget Lab––that overall so far we probably have not seen major impacts of AI on the labor market.

But that said, there are certain segments of the population where we might already be seeing notable impact. And in particular that's for young workers in jobs that are more exposed to AI, such as software development, customer service, things like that.

Eric Brynjolfsson, Ruyu Chen, and I, my colleagues at Stanford wrote a paper called “Canaries in the Coal Mine,” where we documented that employment in these occupations has been declining over the past couple years, and we evaluate some alternatives such as interest rate changes, return from work from home, et cetera.

But there seems to be kind of a robust relationship between these employment changes for entry-level workers and AI exposure.

Kevin Frazier: So it sounds like the question really is about magnitude of these changes and the timing of these changes.

It's no secret that when we see new technology come about, there's always going to be some degree of changes in what professions are rewarded in what way, and how the market has a different demand and supply of those jobs. I wonder, Nathan, for you, when you see headlines like we did later in October of this year, 2025, about Amazon announcing 14,000 layoffs and perhaps as many as 30,000 in the near future, does your mind immediately go to, it's the AI and it's only the AI, we've gotta blame AI for everything?

Or ,what would a economist analysis of that headline look like? What are we missing when we don't get the full interpretation of these sorts of massive disruptions to the labor market?

Nathan Goldschlag: I think it's sort of three things.

The first one is that a company that announces that they're reducing their labor force because of their integration of a new transformative technology sounds a lot better than ‘we hired too many people and we need to contract,’ or, you know, ‘the sales aren't going as well as we thought, we need to contract.’

One of those stories sounds a little better to your shareholders. That's one thing.

Second is that, you know, it's, it might be tens of thousands of employees maybe at one firm, but you’ve got to keep in mind that gross flows in the labor market are enormous, right? So there's millions of jobs that are created and destroyed every quarter, right?

So these, you know, these sort of like drop in a bucket that hits a headline because yeah, the firm dis, you know, firm size distribution is quite skewed, but it's still the case that those sorts of numbers are not going to be moving the needle on the economy. And then the third thing is that in the back of my mind is just as Bharat has said, is that we've, what sort of looked at this question, right?

I've done it. Bharat has done and his team at Stanford, the Yale Budget Lab, like he mentioned. And when you look at these, you know, for evidence of, you know, displacement effects for the economy overall, we just aren't seeing it yet.

So I am open to the idea that there's going to be really concentrated pockets of job displacement, and it may actually happen within certain types of firms, but I don't know that we have the evidence or the foundation yet to make that claim.

Kevin Frazier: So everyone loves to say that AI has a jagged frontier, right? It's everyone's go-to of saying, yes, we will see some displacement in some professions sooner than others, and that's unsurprising, right? When we saw the introduction of the Model T, for example, perhaps the horse pooper-scooper lost their job first rather than X, Y, and Z, you know, horse mechanic or horse keeper.

Man, I am really showing my lack of horse knowledge today, but perhaps the horse pooper scooper was the first to go. Other professions, how they're impacted when and to what extent, is a lot harder to predict.

Anton, I wonder, from your vantage point, how would you characterize the current understanding of AI and its likelihood of causing both short-term and long-term changes to the economy?

Anton Korinek: So the current consensus among economists is that AI is definitely a general-purpose technology, but not necessarily something transformative on the scale of the Industrial Revolution yet. So the current AI systems, they can increase our productivity. Maybe they're going to deliver something like the growth effects of the internet, boom, but they are not going to do something.

That is fundamentally different from that, which is what, for example, lab leaders would be predicting about the economic impact of AI. And you know, the way I view it is, as of right now, the main mode that work interact with AI is the chatbot format.

In a chatbot format, you have turn-taking, you need a human to enter prompts, then the AI can give a response. Then you have the human respond to that again. And that is kind of by design, a technology that is complimentary to work. So we have all these different cores for which jobs are affected by AI, but as long as we use the chatbot to interact with it, it is.

Affected in a complementary fashion. Now, the big change that is currently playing out is that we are moving from chatbot interactions to AI agents, and I personally believe that change is going to also fundamentally change the nature in which AI will affect labor markets.

So the more these autonomous agents roll out, the more AI will become a technology that actually displaces work rather than just complimenting it.

Kevin Frazier: So this really gets at the core concept of the difference between augmenting work and automating work, and the fact that with these AI agents, the presumption or one of the common definitions is an AI agent that can do any task you can do on a computer.

Which for a lot of jobs, if it can perform the email function, the research function, the memo analysis, all of these functions––well, suddenly, for a wide swath of jobs, we are left asking, as you're pointing out, Anton, this isn't just augmentation, this is complete wholesale adoption of the key tasks of that job.

And so Bharat, I want to come back to you for a second, because you flagged that perhaps the tech workers are a little bit more bullish about how soon this automation may occur. Can you give a little bit more detail about why you may be a little skeptical of claims by folks like, I believe it was Dario Amodei forecasting, just, wild end of white collar jobs by 2030, and other folks really warning about the fact that early-stage careers, maybe by the wayside within a matter of years.

What's your own sense of the rate of technology adoption and diffusion right now with respect to that critical question of getting to automation rather than just augmentation?

Bharat Chandar: I think it helps to start by thinking through this historically.

So technologies in the past, they've destroyed work. And at the same time that they've done that, they've created new forms of work, and they've created new labor demand for existing work so that today the unemployment rate is under 5%.

That's pretty low by historical standards, despite all the technological change that we've seen over the past. Couple hundred years that replaced, you know, for example the horse scooper that you were talking about before.

So even as these technologies displace labor, they also create new forms of labor demand and create new forms of work that allow people to continue to find employment in the labor market.

And I think where the uncertainty lies is AI’s fundamentally different in terms of tech compared to prior technologies and I think one way that you could think it might be different is that as, because the models are improving very quickly and because they're becoming more agentic, like Anton was saying, you could imagine a world in which the new work that's created and the new labor demand that's created is also being done by AI in the future.

And I think that is one way that you could think it might be fundamentally different from prior technologies. Now there's a lot of uncertainty about when or if that might happen. And I think a lot of economists would disagree on that possibility.

So I think we're in the stage where we need to do more research to understand how it might impact the labor market going forward, how the model capabilities are improving, and which dimensions that they're improving in.

And is there a possibility that we might be able to even direct the direction the, you know, directs the nature of the technological progress in this technology, so that it becomes more augmentative versus more automated in nature?

Kevin Frazier: Yeah. And I, what I appreciate about this conversation and love talking to economists, because as an undergrad economist, so not nearly on any level like y'all, but what I always loved is we're talking about numbers.

We're trying to quantitatively analyze things. We're trying to do really grounded empirical analysis.

And Nathan, from the vantage point of economists generally, where do we need more data? Where do we need more information? What would make your job easier?

What are some of those kind of big gaps in information that would allow us to have a little bit more certainty about how AI is impacting the labor market and how it may do so in the relatively near future?

Nathan Goldschlag: It's a great question. So I did spend the majority of my career, almost 17 years at the US Census Bureau as an economist studying firm dynamics and AI as well. And so while I was there, we designed new survey questions for firms to figure out who was or wasn't using AI.

One of the things that came out of that work, which sometimes surprises people, is that AI use rates among firms is still like something like eight, you know, nine or 10%.

It's still quite low. It can be really high in the information sector, something like 25%. But overall, the use rates are still pretty low. But in, you know, in terms of the data we would need, I think for, you know, deeper measures of adoption, right, so one of the, one of the questions that we asked in those surveys is sort of, due to the adoption of AI, how did your demand for skill or demand for labor change?

There's lots more we can do with those types of questions to kind of get, you know, self-identified causal estimates from the firm. Because the firm is saying, you know, ‘I use this technology and this is the consequence of that.’ So having larger panels, but then also getting additional data at the worker level.

There's a couple different surveys that went out and tried to ask workers, do you use AI, you know, generative AI in the past two weeks at work or something like that?

You get a much higher, you get a much higher use rate in that case. Something like 25 or 30% or something like that. You know, it's higher than the firm level use rate, but it's not entirely clear what that use is.

So if you're just using it as a substitute for Google, that's not going to generate the sort of productivity effects that economists typically think about. So, additional measures of how firms are using AI, how workers are using AI, and how it's impacting, you know, the demand for labor and skill. I think those are going to be really key.

And the most important bit, right, from an economist's perspective is to have longitudinal data, right? We need to be able to see adoption and de-adoption over time, and to sort of get a sense of how the adopters look different from the other firms that don't adopt, and then how their trajectories change over time.

Kevin Frazier: I love the question of adoption so much because the generic question of ‘are you using AI?’

I mean, sure, yeah. I made a Studio Ghibli meme last week and it was hilarious, and then I was on Sora for five minutes or whatever. Sure, I adopted AI, but to your point, Nathan, that's not the transformative economic use case that we really need to get at.

And what I think is telling as well is we keep seeing these reports. For example, from MIT a few months back, about 95% of all AI institutional adoptions failing. Which, I know, is a study that's very much contested, but it's not asking for example, well, did you do any work to train your employees how to use AI and to actually adopt the tools that were best suited for the task at hand?

So this is just such a complex question. And Anton, I wonder, for you, when you are starting to think about your research agenda, what's top of mind? What are you studying now to try to provide some more clarity around these really weighty questions?

Anton Korinek: I think the work that Bharat and Nathan are doing is really crucial to give us a view of what is going on right now. Or to be honest, if we look at data, it's always a view in the rearview mirror of what has been happening in the recent past, right.

And I think the crucial thing is we have this technology that is rapidly evolving and getting better so quickly, and we need to be prepared. For future scenarios that we can't quite see in the data yet. So in order to get a better sense of what's ahead, I think it's useful to distinguish on the one hand, between the frontier model capabilities.

And then on the other hand, this crucial question of diffusion that Nathan was just speaking about. If we only studied the diffusion with a rearview mirror, it tells us what has happened, let's say, three months ago, six months ago. But it doesn't tell you what we should prepare for in six months from now, or 12 months in the future.

And so observing what is happening at the frontier, what I see is that models are getting better very quickly. I believe that the labor market effects at firms that are using the frontier level capabilities are going to be much starker. Then what we see, let's say for the first three quarters of 2025.

Now there is also a lot of uncertainty about it. That's kind of the downside of trying to predict the future, right? But I think, given that the technology is evolving much more rapidly than any prior technology, it is crucial to kind of embrace that uncertainty and to make sure that we are also prepared for radical scenarios.

So one way of doing that I am engaged in in my work is scenario-planning, trying to predict, what would scenarios in which AI rapidly advances to human-level capabilities across many different areas of application––what would they imply for the workforce, for productivity, and for other economic measures?

Kevin Frazier: Yeah, it's certainly a fascinating thing we're seeing play out across all of these professional domainsm, of anticipating ‘if we had X occur tomorrow, if AGI is reached tomorrow, if superintelligence occurs tomorrow or in the near future, how do we respond?

And that obviously raises a whole can of worms about, okay, if for example, Amazon does lay off all of its software engineers, or a large fraction of them next year or two years from now, that's a lot of smart youth, smart men in particular, just hanging out in Seattle.

What the heck are they going to do? What does that mean for the future of the country? So on and so forth.

But Bharat, I wonder from your perspective and research agenda, what is driving your consideration of helping policymakers respond to this moment? Because one of the things that we've discussed across this trio is over the long term, there's always been a sort of correlation with technological progress and technological adoption, aligning with societal wellbeing and the general welfare improving, all else equal.

The country that leads with technology generally has a stronger economy and more prosperity, but it's not exactly a compelling pol. It's not exactly a compelling political message to get up on your soapbox at that Seattle public market and say, look, Amazon employees, you'll be fine. In 10 years the U.S. is going to be a leader in GDP, so just buckle up and hang on.

So how can you all, as economists, help figure out this very difficult and somewhat unavoidable trade-off, right? We can talk about UBI and some of the other solutions that people may have to this down the road, but in terms of just the historical precedent of saying, ‘we know tech progress leads to some of these trade-offs, how can we navigate them,’ what are your suggestions or what are you thinking through some models for policy makers to keep top of mind?

Bharat Chandar: Great. I think Anton raised a great point just now about if we want to do projections going forward, we can't just look at historical data. We also need economic models to interpret the data and do predictions about what might happen going forward.

And just like him, I'm also interested in doing some scenario-planning around that, particularly on the labor market side. And the nice thing is because we have years and years of research in this space, there are existing modeling tools that we can use to think through potential counterfactual or simulated labor market impacts on different sectors and how that might affect workers going forward.

So that doesn't mean that we're going to see this mass disruption in the next several years, but we can ask questions like, if we did see this, how might that affect the economy in equilibrium? Like, how will people shift across different sectors of the economy, different occupations, et cetera?

And I think an implication of that is we could start thinking through what would be policy responses. There's obviously been a lot of discussion about UBI, but we could probably be more creative about labor market policies as well, and how that might affect the trajectory if we do start seeing a greater job displacement.

So that's definitely on research agenda for me. How do we think about modeling these implications? What could the potential impacts be depending on the level of AI progress that we see?

What are the sectors that we expect to be more impacted? And then how will that reshape how workers, you know, switch jobs across the economy, and where we might expect a greatest growth and then the greatest declines?

Kevin Frazier: And Bharat, just to stick with you for a second, for the early-stage professionals in particular, imagine you've got a 18-year-old daughter.

She's on the precipice of going to Brown or some other-grade school. What is your advice, I'm sure you get this all the time of, do you go into computer science? Do you go into you know, canoe trip planning?

What is the major that you would recommend, or how would you advise just young professionals navigate this turbulent time, especially given your research in the “Canaries in the Coal Mine” paper?

Bharat Chandar: So let me start by just talking as an economist for a little bit.

I do think that we know very little about what's going on the education front, and I think that's kind of crazy. Frankly, I think we don't know how AI is impacting students' choice of major or this choice of career that they want to pursue after school.

We don't know very much about how it's impacting their learning and what are they learning in school, how is it changing curricula, et cetera. So I think there's definitely a lot of need for research in that space.

Now, to directly answer your question, I think the answer that I would give is that these technologies do create a lot of opportunity that we could have never imagined in the past.

I think you can build things, learn about things in ways that were just not possible before. You can learn about a topic, ask an expert any question that you want, and get almost a perfect answer immediately.

And I think that's unbelievable. You can build a website from scratch. So I do think that there are opportunities for building things, learning things that we could have never imagined before.

And it would be, you know, I would definitely encourage young people to make use of those tools.

I didn't have that when I was, you know, entering undergrad.

Kevin Frazier: Okay, so I heard you recommend Canoe Trip as the recommended major.

Bharat Chandar: I'll just, I'll pass. Yeah. Building those social skills. Yeah. Management skills.

Kevin Frazier: There you go. There you go.

So Nathan, you have a unique background with that census portfolio, and the understanding of what are the questions we should be asking, or what are the questions policy makers should have top of mind of tracking? And I wonder when you are thinking about the statistics that tend to dominate the headlines in terms of unemployment rate, or even statistics, as Bharat’s research highlighted, about unemployment among specific groups and adoption among specific communities, what statistics do you think we're not paying enough attention to or trend lines?

You know, what data for policymakers would be the most influential in guiding some of these key decisions over the next few years?

Nathan Goldschlag: So, I think part of Bharat’s answer started to hint at this idea of reallocation, which I think is going to be really important.

So, you know, one of the, one of the things that you know, may come as a surprise to your listeners is that there's been like a, something like a 35-, 40-year decline in business in the United States. And it's not just the United States, by the way. Most developed countries have experienced decline in the rate of new businesses, of people changing jobs, job creation, job destruction.

All those different measures have been declining for years and years after COVID. We had sort of this surprising uptick, right, where there was like a increase in entry, lots of new business formation. Some of that was in response to sort of new opportunities that were, that, like new internet-based businesses, that sort of thing.

But it did sort of. It was a spark of like, maybe this is something that could be sustained and if it was, by the way, would be really important for growth, right? Because new businesses are usually in more likely to introduce new ideas. They play a disproportionate role in job creation.

Now, all of that is to say, right, what does that mean? What does this mean for AI? I think the fir––you know, first things first would be measures of dynamism, right? So a lot of the discussion that we've had so far hints in one way or in another of the, at the reallocation of labor and capital that's going to be induced by this new technology, right?

And so if there are new production processes that firms could adopt that would make them more productive, either augmenting or substituting for labor, all those different things are going to involve some form of reallocation of resources.

If it's the case that certain types of degrees become less valuable because they're, you know, the overlap and the substitution effects are so strong, right, that's going to involve reallocation of degrees as well, where as students enter, to enter college or, you know, post-secondary education, they might be sort of shifting the composition of the types of degrees they get. Same thing could be happening to the occupational distribution.

So I think the statistics that I would say to watch are the ones that are based on reallocation and people making choices based on a changing landscape that they're facing, changing incentives.

And the same thing, by the way, is, it would be my answer for the 18-year-old that's thinking about what degree. So it's sort of like you, you need to have an eye on the ways in which whatever you choose to do is going to be impacted by AI and the ways that it can potentially improve the things that you do, but then also a heavy emphasis on communication and interpersonal communication skills and being able to sort of lean into those soft skills that are going to likely remain sort of something that humans are better at.

Kevin Frazier: And this reallocation concept in dynamism more generally, I think is so important to call more attention to, because folks just aren't used to talking about these terms of saying, well, a healthy economy isn't necessarily one in which you have the same job for the entirety of your career at the same company.

In fact, it's arguably way better on the aggregate to have as much entry and exit of firms, as well as your ability as a worker to move across state lines to move into a new profession, so on and so forth.

And yet there's a lot of stickiness, both with respect to the firms we have in the market currently, and with respect to ourselves as being able to move to a new state.

As someone who's moved to, God, what does my wife remind me of? I think it's seven states that we've moved to, it sucks. I hate it. I can't tell her how much I hate it or else we'll never get to move again. But no one likes moving.

But if you want to have that economic dynamism, especially in a time where we're going to have new jobs get created for new markets that no one can predict, that's all the more important.

And so, Anton, I know the temptation when talking to a group as smart as you all is to say, please solve this. What is the bullet, the silver bullet solution to alleviating everyone's concerns about the AI economy? But I'm going to flip that on its head because I'm a nice host, and instead, I would like you to identify one or two of the worst ideas you've heard.

What is a solution that you, a solution in air quotes, what is one of the solutions that's been offered that you just think, oh my gosh, if we follow that, ugh, bad things ahead. What is one of those for you that stands out and why?

Anton Korinek: Well, Kevin, if you allow me, I will start by taking a bit of a step back.

Kevin Frazier: Go, go for it.

Anton Korinek: The ultimate objective, both in our economic models and what I suppose all of us economists are striving for is human welfare, not GDP.

In many times those two kind of move in tandem. People generally don't like recessions, which is when GDP goes down. People love booms when GDP goes up and all boats are lifted. But sometimes the relationship does not hold one for one.

And especially if this technology turns out to be really labor-displacing, then it may be one of those episodes. So coming back to what are bad ideas? I will say this a little bit tongue in cheek because you focus so much on it, and it depends very much on which scenario we are going head into the future, but so there is this possibility out there our leading labs are going to create something like AGI, artificial general intelligence,. Which by their definition, if you look at, let's say OpenAI's charter would imply AI systems that can effectively accomplish most economically valuable jobs. So if they do reach that, and right now we have to all acknowledge there's a tremendous degree of uncertainty about it.

And it's speculation. It's no certainty at all. But if that happens, then at some level labor, which would now have to compete with these machines that can do essentially most economically useful tasks would be fundamentally devalued. And then just prescribing more dynamism, as you just did, may be a very bad idea. So in the short term, while labor is extremely valuable. I agree that may be a good prescription.

Although we should also emphasize that people generally like some stability, right? So we don't want just unfettered dynamism, everybody has to change their job every day. Because stability is also something valuable in people's lives, especially if you try to raise a family, for example.

Now, but if our economy fundamentally changes. If it becomes an AGI-driven economy, then I think we have to acknowledge that the role of labor will have to decline and just telling people to be more dynamic would not be a solution.

Kevin Frazier: So Bharat, I'll turn the question to you. The worst or among the worst ideas you've heard, and I will recognize Anton's wonderfully phrased caveat of course, the worst solutions today might actually be some of the best solutions in a different future.

But in terms of any ideas you've heard so far that you just want to yell from the rooftops, ‘please no,’ what comes to mind for you? Because I think that your very practical research has received a lot of headlines and a lot of attention from people who are interested in a lot of these policy conversations.

Bharat Chandar: Well, I think one side of that is that I actually haven't seen very many serious policy proposals for dealing with AI, and that might be because we're in the very early days of studying it. I do think, you know, we've heard a lot about UBI, which is, you know, not the most creative solution in the world. And I do think, you know, we value work for reasons other than just money.

You know, I think it's a useful thing to consider, but we can also be more creative about policies that we think about. I do think one idea that this isn't just relegated to AI, but I think the de-growth idea is probably not a very good one. So the idea that we should stop economic growth or even scale back our you know, technology or material wellbeing because we're worried about impacts in various domains, including potentially AI, I don't think that's a very good one.

I think there're––if you're worried about certain aspects of the technology, there are policy proposals that you could consider that might mitigate those without compromising the wellbeing and the growth in material consumption for future generations in a way that could be very harmful.

So that's probably one that I would mention.

Kevin Frazier: Yeah, and that really builds well on Anton's point of while GDP growth and economic welfare may not always be one-to-one, certainly GDP growth, declining and economic welfare usually are not aligning very well. So thanks for sharing that.

And Nathan, no pressure, but I expect a home run from you.

You've had the longest talk, all right, about this issue. So blow me away. What––

Nathan Goldschlag: So it’s not, it's not fair. It's not fair because Bharat took the easy one that is low hanging fruit. Degrowth is a terrible idea. Okay. Alright. So that's the easy one.

I'd say there's a, there's an interesting tension in this discussion, which, you know, I would say it's not necessarily you know, some specific policy idea, but there's this, you know, idea that kind of floats around the background of folks that are really concerned about AI that’s, ‘can we freeze the economy in amber, right?’

There's a lot of, you know, sense that, well, the occupation distribution we have now is somehow a sacred cow that we need to make sure persists, right? And I think that's a mistake, right? I think David Autor has this nice summary paper of like a hundred years of automation research and, alright, so something like 40% of employment was in agriculture in 1900, and by 2000 it was like 2% right?

And so, you know, the concerns about like the total substitution of labor, they, that sort of makes sense. But you know, if we're sort of in the normal realm of the evolution of technological change, and maybe it's a little faster, maybe it's a little more intense, there's more occupation mix, switching, like all that sort of things.

So long as we're within the normal bounds of what we've experienced in the history of technological change, right, I think you know, the––you'll want to facilitate the reallocation rather than trying to freeze the current rea-, the current composition of the economy in amber.

Kevin Frazier: Yeah. And I love that because. I wasn't being facetious, or totally facetious, when I would brought up the horse pooper scooper.

I went back to the New York Times archives and tried to find any coverage of, you know, the protests of the horse pooper scooper, or let's protect the horse pooper scooper. You don't find it. There's not a, you know, we need to make sure that we look after the livelihood of this specific occupation.

Should we look after the livelihood of those individuals? Of course. But this is a really complex question, so with that said, one final kind of lightning around it, and I'll just defer to anyone who wants to raise their hand and suggest an answer here, which would be, is there a jurisdiction today that you think is approaching the AI and economy question well?

And this can either be from a data-gathering standpoint or from a policy adoption perspective. Is there a jurisdiction who you would say, more folks should know about how X is handling AI and the economy?

Nathan Goldschlag: I'll take a shot at this, you know, shameless promotion of the Census Bureau.

Kevin Frazier: Love it.

Nathan Goldschlag: I think you know the history of measuring technological change at, you know, through the federal use of federal statistics.

It very often is the case that the stats agencies start measuring it after it's basically already fully diffused. Right. There, you know, there's not a lot of cases that you'll find where the federal statistics were sort of front looking around the corner at what's to come. And I think with AI and robotics it, AI, the Census Bureau actually caught it, right, so there was questions that were asked in the 2018, or 2017, or 2018 ABS, annual business survey.

So they've been asking questions about AI use since before ChatGPT hit the scene. And so that actually lays the foundation for economists to really understand, measurement––measure, understand the measurement of the effects of AI in a much more robust way, having that longitudinal, sort of, before the release of those technologies and then afterwards.

So I think that in the case of federal statistics, we have sort of a unique opportunity in the measurement of AI to sort of see how these things are affecting firms and individuals in real time.

Kevin Frazier: Wonderful answer. And I will accept the shameless self-promotion. That's always welcome among academics. Anton, Bharat, any, anything coming to mind? If not, that's okay.

Bharat Chandar: Alright, Anton's going to pass. I do think I, I have a quick answer here. I do think that we benefit from the fact that the research in this space is often made publicly available. And that's true both at academic institutions but also at the lab.

So there's definitely a lot more data that we could ask the labs to report, but I do think it's kind of great that they, a lot of the research that's going into the production of these frontier models, that's being made public.

A good example of this is GDP valve from OpenAI, which is kind of documenting improvement across tasks specific to occupations over time. And you know, they're kind of tracking that the tasks that are, that apply to different occupations, the models are improving pretty quickly in their ability to do those.

And I think that's quite useful. Anthropic is putting out the Anthropic Economic Index, which gives us a sense about how different occupations are using AI. And of course there's a lot of research coming out of the academic community as well, documenting some of these facts.

So I do think it's great that we have some insight into this because the research is fast-developing, and it's also kind of public, and something that we can look at as economists or other scientists who want to study the capabilities and other changing, you know.

That said, we can definitely ask them for more data in different domains. I think we all agree that we need to be doing better data collection, but I do appreciate that this is a space where there's serious scholarly query and public dissemination.

Kevin Frazier: All right, Anton, we'll give you the final word.

Anton Korinek: Alright, I'll add one more quick point which builds on what we've already discussed before.

So I think economists are kind of inherently uncomfortable making too many predictions, and especially if this changes super quickly. And that's why I want to reinforce this need for scenario-planning.

And I think it's so important, even though we are uncomfortable with, even though we have to make lots of assumptions and, of course, most of our predictions are going to turn out to be false. I think it is crucial to engage in scenario-planning to head into this rapidly changing future.

Kevin Frazier: Well, I had a scenario plan in my head of when three economists walk into a podcast, what the heck happens?

And I had to admit most of those futures, I thought, ah, this may be a little boring. This was awesome. You all rock.

I will look forward to the day of having you all back on to share your next research, but thank you all again for taking the time.

Anton Korinek: Thank you so much.

Bharat Chandar: Thank you.

Nathan Goldschlag: Great to be with you.

Kevin Frazier: Scaling Laws is a joint production of Lawfare and the University of Texas School of Law. You can get an ad-free version of this and other Lawfare podcasts by becoming a material subscriber at our website, lawfaremedia.org/support. You'll also get access to special events and other content available only to our supporters.

Please rate and review us wherever you get your podcasts. Check out our written work at lawfaremedia.org. You can also follow us on X and Bluesky.

This podcast was edited by Noam Osband of Goat Rodeo. Our music is from ALIBI. As always, thanks for listening.