AI, Financial Markets, and the Economy

Itay Goldstein: Welcome, everyone. This is the third season of the Wharton “Future of Finance” series. I am Itay Goldstein, a professor in the finance department, currently the chair of the finance department. In the third season, we are talking about AI in finance.

One of the most exciting developments we have in recent times is AI. And this is affecting every aspect of finance and the economy. Today we are going to try to understand at a bit of a broader level what is the effect of AI on the economy, how this is reflected in financial markets, what can we learn from financial markets, and what are some of the risks we might have in financial markets because of a potential bubble.

To dive into these issues, we have two perfect guests. First, we have Maryam Farboodi. Maryam has been in different universities in the last decade or so, starting in Princeton, then moving to MIT and now is new at Cornell, where she is joining as an associate professor in a joint appointment between the business school and the computer science department. Hi, Maryam.

Maryam Farboodi: Hello. Thanks so much, Itay, for the very kind introduction.

Goldstein: Thank you. It’s great to have you. And then we have my colleague here at Penn over in the economics department, Jesús Fernández-Villaverde, who is the Howard Marks Presidential Professor of Economics. He has been working over the years on several related topics, including economic history. We can learn a lot from history when thinking about AI and also computational economics. And certainly there is a lot here that has to do with computation and programming.

I’m going to start with you, Jesús, given your broad historical perspective. Can you help us understand what is new in the current AI wave and how this compares to previous technological revolutions?

Jesús Fernández-Villaverde: Thank you for having me here, Itay. I like to think about artificial intelligence as a general-purpose technology. This is a very common term in economic history. It refers to technologies that are a very broad spectrum that affect everything in life. So think about electricity. Electricity is just not about the light at your home, but it’s also the fact that you can reorganize factories around the electric engine, that you can completely reshape society. And before that, perhaps it was the steam engine.

So around every 100 years or so, you have one of these general-purpose technologies coming that completely change your life in nearly every aspect of it. And I think that artificial intelligence is one of those. This is really incredibly exciting, because perhaps the internet was also a general-purpose technology, but I don’t think it was as much as artificial intelligence. I really think we are in one of the situations that happens once every 100 years where the whole economy is going to be completely reshaped.

Goldstein: Very good. Maryam, you recently wrote a paper that is trying to understand, based on information in financial markets, what the market thinks about this technology and how it’s going to affect future growth or maybe even some catastrophic scenarios like extinction. Can you help us understand based on your paper — first, what motivated you, what you looked at, and what are your conclusions from that?

Farboodi: It’s so exciting for me that you brought up this question. Because in this paper, we asked a super simple question: Do markets believe in transformative AI? And that includes all the aspects that you said. Like some extreme event, if you want. Extreme good, extreme bad, or different rates of future growth rates. We approached this question by looking at the government bond market, given that market prices have long been, way before AI, used as a tool and are a strong indicator to understand investor beliefs.

To be very honest, when we entered this, we did not expect to find anything, given that we did not think the U.S. government bonds would be impacted by such a rapidly moving technology, which is so young. But to our extreme surprise, and contrary to our initial intuition, we found strong evidence that the investors do believe in transformative AI. And in particular, we find that the long-term treasury yield fell economically and statistically, and statistically significantly, around major generative AI and new model announcements, since the announcement of ChatGPT 3.5, which got attention of all the market participants to this new wave of technological growth.

Then, given that this finding of drop in yields — and such significant, large drop in yields — was very counter-intuitive to us, because it would go against the observation that investors believe in any extreme events, and would also go against the observation that they would think, they believe, in broad-based higher growth rates. We decided to think about this yield response [as it relates] to investors’ beliefs about actual technological growth rates — whether AI’s technological growth rate is going to be faster or slower. And in order to do that, we used survey data from online platforms, which got participants’ beliefs about the rate of growth of AI technologies when these new models came out.

Again, to our second extreme surprise, we found that these declines in yields are very concentrated among the announcements that are perceived by market participants as good news about AI progress, while in the announcements that were not perceived as particularly good news in terms of AI growth, the response was way attenuated and not statistically significant at all.

Goldstein: So what’s your take on it overall, if you had to summarize?

Farboodi: This is a very, very good question. That’s a million-dollar question. From finance, we know that U.S. government bond yields, to the extent that they’re safe, reflect how investors discount their future payoffs. So if bond yields are falling, that means investors have a higher discount rate. One way to read that would be that the investors’ expectation about the growth rate of the economy is going to be lower. [That] sounds extremely counterintuitive, because these drops were exactly concentrated in events which were [related to] positive AI growth.

So to be perfectly honest, we cannot pinpoint a specific answer. But there are a couple of mechanisms that we have explored that are broadly consistent with the drop with good news. One is that if people believe that gains from AI are very concentrated within a part of the population, [and that] this growth in AI would have adverse wealth distribution away from the average population. And so they would be concerned, and that would push down the yields.

Alternatively, a saving cut story, where all the rich people are going to get richer, then they want to save a lot, and that would push down the bond yields — that’s another reason. An alternative reason that some people have suggested is that the progress of AI improves the U.S. fiscal position, and that brings down the probability of defaults of U.S. bonds. This is the only thing that we can actually directly test by using the CDS (credit default swap) spread, and we do not find any evidence about it.

So to us, something about the concern about redistribution or saving cut seems to be the most consistent with what we have seen. But a lot more work is needed by people like you to figure out the actual mechanisms.

Goldstein: Yes, very interesting, just looking at what is the reaction to the new release of new versions of the product, and also just looking at these rates. So, Jesús, are you worried about some of these things if you think about redistribution, and what that is going to do overall to inequality and distribution in the society?

Fernández-Villaverde: I tend to be relatively optimistic that, contrary to some expectations, artificial intelligence can have a positive effect on the distribution. If we go to the United States and other advanced economies, most of the economic growth from around 1973 or so on until around 2010-2012 was what we economists call a “skill bias.” Basically, economic growth complemented very well with skill. So if you look at the relative wages of people that go to college versus high school graduates, people that went to college did comparatively much, much better over the last 50 years.

And the reason why that happened probably had to do with the type of technologies that were implemented in the 1970s, 1980s, and 1990s. Computers, etc., which complemented very well with skill. For the first time in over half a century, we have a technology that substitutes high skill, but complements, at least for the moment, the type of things that are hard for machines to do, but that humans have a comparative advantage.

So what do I have in mind? It turns out to be the case that right now you can use Claude to write a contract. You can use Claude to deal with low-level legal cases, and you don’t need a lawyer anymore. I actually got involved — just as a personal anecdote, last summer I got a speeding ticket in Germany. It was completely unfair, so I actually appealed [against] it. And I appealed in German, which I don’t speak that well, and I certainly cannot write a legal document in German, through Claude, and I actually won the case. You know, 20 years ago, 10 years ago, that would have required me to hire a German lawyer.

On the other hand, artificial intelligence cannot do anything about the landscaping in my house here in Haverford, PA. So that means that over the next 20 years or so, the relative wages of landscapers is going to go up, and the relative wages of lawyers is going to go down.

Now, it is going to be the case also that those that own shares of some of these companies may make a lot of money. Now, this may happen. I tend to be on the side of those who think that’s not going to matter that much, because I think that competition is going to push the price down. And we saw it recently with Fable, the new model of Claude. Claude tried to keep Fable only as an extra tier. They realized that after one month they cannot really do it. I was using Fable half an hour before, with my regular subscription.

So in that sense, yes, there are going to be some people who are going to make a lot of money. But by and large, my baseline scenario is that in 25 years, 30 years, we are going to look at artificial intelligence on the whole as something that has not increased income inequality within society. What worries me much more than increases in income inequality is that vested interests may actually prevent the rollover of artificial intelligence in many sectors. Again, economic history tells us that just because we have a new technology doesn’t imply we are going to use that technology.

My favorite example always is about the printing press, which was the artificial intelligence of the time. Suddenly we can make a book, we can share information. Well, the printing press was prohibited in many countries. The printing press was extremely limited, for instance, in the Ottoman Empire for several centuries. And the Ottoman Empire didn’t take advantage of the printing press.

So what I’m afraid of is that we are going to see, in many countries, roadblocks to the implementation of artificial intelligence, just because the vested interests are against it. And that’s the place where maybe thinking about redistribution that gets around those roadblocks may help everyone.

Goldstein: You think that those roadblocks are generally bad? You don’t see any positive aspect of them?

Fernández-Villaverde: I’m not going to say all of them are bad. Like everything, you are going to find exceptions. But I think that nine of each 10 limits that I can see regulatory agencies or parliaments across the world passing against artificial intelligence are not going to be a very good idea. Anyone that drives in New Jersey wonders why in the world in New Jersey I cannot pump gas myself, despite the fact that in the rest of the planet everyone pumps their own gas, and nothing happens. So ’those are the type of problems I can see happening. [Such as] saying, “Oh, you cannot get any type of health service through artificial intelligence.”

Just to wrap up, let me give you a very concrete example. In the European Union, it’s extremely difficult to create any education-related artificial intelligence, because there is a regulation in the European Union that says that if you are going to be dealing with children or with the education service, you need to do so much regulatory compliance that it makes it nearly impossible for any startup to have the personnel to be able to comply with it. So you may end up in a situation where China has fantastic artificial education apps, the U.S. has fantastic apps, [but] Europe just doesn’t. Good luck trying to convince the vested interests in Europe that this is just a silly idea.

Goldstein: Yes. Maryam, how do you think about those tensions — the societal and political tensions, and how they interact with the economic effects of AI?

Farboodi: I don’t disagree with Jesús that it is important that a regulation is set to allow progress of AI. And this spills over to many, many different industries, effectively. At the same time, I’m particularly concerned about education. I agree that jobs like [those of] miners or landscapers would not be impacted. But there is going to be a lot — a broad set of not uneducated but lower education people — who can be completely replaced.

Unless there is — and it’s possible to have a good education system — an education regulatory framework to allow these people to train better and learn the skills to augment their abilities with this new technology, I am concerned that this kind of, if you want, middle part of the population, which is actually large, would be very adversely impacted and would not benefit from the growth rate of AI, even if the aggregate GDP growth is positive.

Aat least in the U.S., which is at the forefront of this technological growth, there does not seem to be a lot of discussion or investment in how to retrain individuals to be able to best use AI. I think that’s a concern.

For instance, going back to Jesús’s very, very interesting example of him winning a law case. This is Jesús He is a professor at UPenn, okay? And of course, he has the skill sets to use this technology as good as anybody can. But [as for] the average population, I am not sure they’re benefiting from this advance as much as they should and as much as the potential allows.

Goldstein: Very good. So, let me move on a little bit and talk about what is going on right now in the stock market and the way that this is reflecting the overall investment and potential of AI. Of course, we saw a lot of money being spent on AI investment across the board, and we see valuations at close to peak level. There is this constant worry that maybe we are facing a bubble and if the bubble bursts, that is going to have very negative implications, of course, for everyone. And constantly we hear about comparisons to previous episodes, most recently the dot-com bubble. So, Jesús, when you look at all this, what is your take? I know that you are a little more skeptical.

Fernández-Villaverde: Yes. So, first of all, I have a little bit of a heterodox reading of the 2001 episode. Yes, we had a big increase in valuations, and the market corrected in 2001. But if you actually look at what the internet delivered by 2015-2016, it really did deliver what those valuations in 2001 were suggesting, and perhaps even more. It is true that in the short run it was a bumpy ride, and we have seen that with any new technology arriving in the U.S. since the railroads. And it is true that some of the concrete companies that had higher valuation in 2001 are no longer with us. But the amount of wealth and value generated by the Amazons of the world, by the Googles of the world, by the methods of the world has been just absolutely incredible. So, I think that in retrospective, maybe we can go back and read that 2001 a little bit more positively that what we gave it credit for.

Now, is this time going to be different or not? First of all, I’m not claiming that we may not have a crash on the market. This very well may happen. I’m just claiming that when we go back, in 2045-2050, [to] this moment, the preponderance of the evidence is that we are going to say, “Yes, artificial intelligence was activated.”

Now, what I think makes the case of 2026 a little bit different than the case of 2001 is the nature of the technology. In 2001 we were raising a lot of money to buy servers. To buy, or  to put fiber cable around the country to basically generate infrastructure for the internet. That infrastructure has been around for 20 years. It is still useful.

The problem is, if I spend $5 billion dollars or whatever it needs to spend right now to train the most recent version of a model, as soon as the new model comes, the depreciation of those $5 billion dollars is close to 100%. And that tells you that the position of these companies may be much more fragile in the long run than some of the internet-related companies. In particular, the guys who may end up being the dominating actors in 2035 may ’not be anyone who is even alive right now. And we may end up [where] a lot of the money that we put on OpenAI or that we put on Anthropic, 10 years down the road, doesn’t deliver a lot of concrete return to the shareholders of that concrete company. So it’s a little bit of a mix of, yes, [but] on the whole, I think this is going to be a big deal. In particular who knows, is Anthropic going to be around in the years [ahead]? I don’t know.

Goldstein: Yeah, so you’re more optimistic about the longer term, but you don’t really know what’s going to happen in between [now and then].

Fernández-Villaverde: If I knew that, I will be trading on it. And that house and landscaping is not going to pay for itself. No, I’m very optimistic. I really think that we are going to be back in the year 2050 looking at what has happened over the last 10 years, and say that was an important inflection point in the economic history of humanity.

Goldstein: But just to be clear, one difference between the current technology and previous technologies is the substitutions with humans, right? And so the counter argument to everyone who is optimistic is always, “Yeah, but, you know, at the end of the day, this technology will do everything. So how can humans be better off? Or what will they do?” So how do you see that being resolved?

Fernández-Villaverde: Yes, so if you actually look again at the historical record, we have been saying that for 250 years and it never happened. Watch Metropolis by Fritz Lang (a 1927 science-fiction silent film). People were saying [back then], that everything was going to be done by robots by now. It still hasn’t happened.

Now, in the case that artificial intelligence becomes so incredibly good that we don’t really need humans to produce cars or to produce a lot of other services, I think that we are extremely creative as a species to come up with other ways to fill our life. Okay? And there’s a lot of things I would like to do with my life if someone walks in and says, “There. You have, Jesús, $200 million, do whatever you want with your life. You don’t need to work for the rest of my life.” I promise you, I will be able to fill every single day of my life until the day I die.

So if this is the case for humanity as a whole, yes, there is going to be an issue about the meaning of life. And as a society, we will need to tackle it. But ’I’m not particularly pessimistic about it.

On the other hand, my baseline scenario is we are just going to find other ways to do things. We are just going to create new products that are going to be complementary to what humans do. And at the end of the day, the result is going to be okay.

Goldstein: Okay, that’s a very good way to think about things. Maryam, you talked before about the paper that you have on bond prices and what you’re trying to infer from changes in the rate after the release of new versions of AI. But over the years, you’ve done research also on the stock market and information in prices. I’m sure that you are looking to some extent at stock prices right now. What do you make of that?

Farboodi: This is very interesting in the sense that there are two patterns happening at the same time. One, in the past two years, everybody is like, “Oh, my gosh, stock prices of AI-related firms are shooting to infinity.” So their valuations are extremely high. Like the NVIDIAs of the world. In fact, there’s an ETF which follows AI-related stock prices.

’[That] two years, two-and-a-half year, or three-and-a-half year rise in stock prices is undebatable. At the same time, when we use the same methodology to look at the response of, let’s say, both the S&P 500 and individual companies like NVIDIA to these new AI model releases, we do not see a large impact. We do not see a statistically significant movement either up or down. Okay. So that is very interesting.

One thing that our results merging the findings in bonds and stocks points to is that to the extent that we believe our bond market results and there is a discount-rate effect, especially for the long-duration bonds like 20 years, 30 years, this is also supposed to show up in stock prices. And particularly AI-related firms, because their cash flows are pretty backloaded, and their valuation would be very much impacted by this. So this kind of discount factor effect would act like we have a bubbly effect on them.

In fact, we did try this with the ETF that I told you [about which] follows AI firms. We do exactly see that that particular ETF, after residualizing on all the other market components, does go up at the time of these AI news events. That is completely consistent with just a bubble discount factor without letting us disentangle that from an actual cash flow factor, which would say that AI would actually make these companies more productive.

That is something that we still don’t know how to separate from each other. And that is actually why we try to, in the first step, look into the bond markets to look at that belief factor before going into the actual cash flow increase due to the advent of AI.

From all that, I could see we cannot rule out a bubbly component in stock prices. I also think, related to the dot-com bubble boom, let’s say, there was a lot of “winner takes it all” in that boom. Maybe, to some extent guided by that, investors now are throwing a lot of funding into the leaders of this new technological revolution, and hope to be part of the winners. Which would not be irrational, but that can rely on this speculative view of the prospects of the AI.

For instance, ’as Jesús pointed out, the cost of this technology seems to be extremely high. So it is unclear if this level of growth is sustainable.

Goldstein: Okay. I guess both of you are also, of course, teaching in a university and constantly interacting with young students who are looking to go into the job market and thinking what they should do. What do you tell them these days, Jesús? Let’s start with you.

Fernández-Villaverde: I wish I knew how to tell them. No. Okay. I actually have been thinking a little bit about it. If any of the listeners follows my X feed, I try to post on that from time to time. There are two fundamental skills that in some sense were already the skills of classical education, and we need to go back to them. The first skill is to know how to ask questions. So think about all the dialogues of Plato. It’s all about how to ask the right question. And the Greeks understood that, the Romans understood that, and maybe we have lost a little bit of that understanding over the last 200 years. So the universities need to be much more about what are the right questions to ask.

The second fundamental pillar of a modern education should be, how can I tell this answer is good? So you are going to ask Claude, “Tell me about X.” Claude is going to give you a very good answer. No. It’s going to give you something that looks like a very good answer. And you need to tell if this is the right answer or not. So if I were the president of Penn, which I will never be, or the dean at Cornell, which I will never be, my whole point will be, “How do I reorganize education around these two points?”

Now, if you are just a student and you are asking that question for yourself, that’s maybe what you can try to get out of your courses. How do I learn what are the right questions to ask? And how do I learn how to judge that an answer that I get from the question is the right one? For instance, in my own courses at Penn, that’s what I try to organize my courses around. Trying to say, “Look, I give you a paper.” This is even for undergrads. “Are you convinced by the answer of that paper or not? Do you think that the answer is the right one? And if the answer is yes, why? And if the answer is no, why?”

And I keep pushing them, “Don’t try to pander to my beliefs, because I’m not going to tell you what I think about this paper. What I want you to think carefully about is: how do you structure an answer and how do you evaluate it?”

Goldstein: Okay, Maryam?

Farboodi: I love Jesús’s answer about education. A lot of it is about questioning and curiosity and validation. And that is what I try to tell my students, that it is important for them to learn. They need to learn how to augment their skills by AI and not to be able to be replaced by it. So one practical thing that I do is that there are a lot of MBAs. I give them questions which are totally mathematically correct. They can give it to AI and get the answer, which is completely right. But there is something that economically doesn’t make sense.  I always ask them, “What is it in this question that does not economically make sense? What part of the input do you think has to be changed to address that economic inconsistency?”

I’m sure an LLM would be perfect at answering this question in six months by the next time that I feed. So far, at least in the past years, they have not been good. And that helps students exactly do what Jesús was mentioning, hopefully.