Matthew O. Jackson is perhaps the world’s most renowned scholar of the economics of networks; as a 2005-06 CASBS fellow, he wrote most of his still-influential book Social and Economic Networks. In this wide-ranging conversation with 2025-26 CASBS fellow Rajiv Sethi, Jackson discusses his foundational work on strategic modeling of networks, empirical applications on the role of economic connectedness in influencing people’s life trajectories in the U.S., related multi-disciplinary and cross-national work he is undertaking at the Santa Fe Institute, and recent cutting-edge work using large language models to gain insights into human motivations and behaviors.
Matthew O. Jackson is perhaps the world’s most renowned scholar of the economics of networks; as a 2005-06 CASBS fellow, he wrote most of his still-influential book Social and Economic Networks. In this wide-ranging conversation with 2025-26 CASBS fellow Rajiv Sethi, Jackson discusses his foundational work on strategic modeling of networks, empirical applications on the role of economic connectedness in influencing people’s life trajectories in the U.S., related multi-disciplinary and cross-national work he is undertaking at the Santa Fe Institute, and recent cutting-edge work using large language models to gain insights into human motivations and behaviors.
Matthew O. Jackson: Stanford faculty page | Personal website | CASBS page | Wikipedia page | Google Scholar page | National Academy of Sciences bio | Stanford profile | SFI page | NBER working papers | Jackson CV |
Rajiv Sethi: Barnard faculty page | Columbia page | CASBS page | Google Scholar page | SFI page | Rajiv's Substack newsletter, Imperfect Information |
Matt Jackson works referenced in this episode:
Matthew Jackson and Asher Wolinsky, "A Strategic Model of Social and Economic Networks," Journal of Economic Theory (1996)
Matthew Jackson and Alison Watts, "The Evolution of Social and Economic Networks," Journal of Economic Theory (2002)
Raj Chetty, Matthew Jackson, et al., "Social Capital I: Measurement and Associations with Economic Mobiliity," Nature (2022)
Raj Chetty, Matthew Jackson, et al., "Social Capital II: Determinants of Economic Connectedness," Nature (2022)
Chetty, Jackson, et al., Opportunity Insights Social Capital Atlas (website)
Dynamics of Wealth Inequality project (Santa Fe Institute)
Matthew Jackson, Social and Economic Networks, Princeton University Press (2008)
Matthew Jackson, The Human Network, Penguin Random House (2020)
Mei, Yuan, and Jackson, "A Turing Test of Whether AI Chatbots are Behaviorally Similar to Humans," PNAS (2024)
Xie, Mei, Yuan, and Jackson, "Using Large Language Models to Categorize Strategic Situations and Decipher Motivations Behind Human Behaviors," PNAS (2025)
---
Rajiv Sethi's latest op-ed is "Polymarket Anonymity Must End," Financial Times (May 7, 2026)
Subscribe to Rajiv's Substack newsletter, Imperfect Information
Narrator: From the Center for Advanced Study in the Behavioral Sciences at Stanford University, this is Human Centered. For nearly a century, sociologists and anthropologists have studied how people organize through networks. And how this relates to human behavior, interaction outcomes, and social mobility. In the last 30 years, economics has jumped into the ring and its tools and methods have transformed the field. Among the founders and perhaps the most renowned scholar associated with the economics of networks is Matthew O. Jackson. Today on Human Centered, a conversation with Matt Jackson, the William D. Eberle Professor of Economics at Stanford University. With expertise encompassing game theory and microeconomic theory, In his most celebrated work, Jackson has analyzed the formation of networks and how they mediate the structure of social relationships with empirical applications in a variety of domains. In recognition of his advancement of network science, he's been elected a member of the National Academy of Sciences and the American Academy of Arts and Sciences, as well as a fellow of the American Association for the Advancement of Science. He's received a Guggenheim Fellowship and received such honors as the BBVA Foundation Frontiers of Knowledge Award, the Jean-Jacques Lafont Prize, the Social Choice and Welfare Prize, and the B.E. Press Arrow Prize for senior economists. And notably for our purposes, Jackson was a CASBS fellow in 2005-6. It was an incredibly fertile year for him as he wrote most of his still influential book, Social and Economic Networks, published in 2008, and worked on at least 12 journal articles. Joining Matt Jackson in conversation is Rajiv Sethi, a professor of economics at Barnard College, Columbia University, and a 2025-26 CASBS fellow. Currently, Rajiv is pursuing two book projects leveraging his deep expertise, one on the interpretation of signals and the other on prediction markets. He's written about both in his popular Substack newsletter. We'll provide links to it and other works in the episode notes. As you're about to hear, Rajiv engages Matt on foundational work on the economics of networks, including Matt's most cited article, a 1996 co-authored paper on a strategic model of social and economic networks that introduced the concept of pairwise stability. They move on to empirical applications of network analysis, including ongoing work with Raj Chetty on the importance of economic correctness as a form of social capital in influencing people's life trajectories, as well as multidisciplinary multinational work Matt is undertaking with collaborators at the Santa Fe Institute. Of course, AI is on the radar of a lot of scholars these days, so we'll hear the two delve into some of their recent explorations and perspectives on large language models as it relates to networks and the social construction of knowledge. Let's listen to Matt and Rajiv think it through together.
Rajiv Sethi: Hi, Matt. Thanks for joining us here at CASBS for a conversation.
Matthew Jackson: Thanks, Rajiv. It's wonderful to be up here on the hill again. It's always so idyllic.
Rajiv Sethi: Yeah, you spent a year here a while back.
Matthew Jackson: Yes, I was here, I guess, in 2005, 2006. And yeah, it's one of those memorable years that sort of deepen your memory.
Rajiv Sethi: Yeah, I'm about halfway through my fellowship here and it's going too fast.
Matthew Jackson: Yes, exactly.
Rajiv Sethi: So Matt, maybe we could start talking about networks because I think You've done more than any other economist to bring this into focus as an object of interest for social science more broadly and economics in particular. But networks, of course, have been looked at in many other disciplines, sociology in particular, I think maybe anthropology also. And maybe we can start by talking about what economics brings to the study of networks that is novel and insightful.
Matthew Jackson: Yeah, I mean, I— so I first got interested in networks through a I guess, serendipitous lunch conversation with Asher Wilinsky at Northwestern. And we were just talking about power, you know, how do people get power? What does it mean to be powerful and being connected and how would you actually measure that? And at that time, this was the early '90s, the sociology literature had been at it for about a century. So they'd been looking quite deeply into this. And I think what economics, you know, an economist's view was different than a sociologist's view of the question. So we were very interested in incentives. We were very interested in, you know, who chooses to be friends with whom. How does somebody like a, um, a Cosimo de' Medici build their network? What are the incentives to build the network? And, and so I think economics brings a couple of things to the table. One is, um, a choice-based perspective, and the second is a welfarist approach, right? So, so we start thinking about utilities and preferences, and you can evaluate a network as being socially better than another network in terms of how it manages to communicate information, how it aggregates information. So the perspective of an economist was slightly different. So it sort of complements the sociology literature was already there. And then it brings a lot of different applications to the table as well: financial networks, supply chains, countries and trade patterns. So there's a lot of other things that end up being on an economist's plate that are natural to look at.
Rajiv Sethi: Yeah, that's really helpful. I wanted to ask you also within economics, methodologically, that paper with Wolinsky has been very influential. I think it's your most cited paper and you introduced the concept there of pairwise stability. For the economists who are listening, Did you see this as a departure from methodological individualism to some extent? There's a sort of mixture. There's a sort of cooperative that's not entirely non-cooperative because there's a process of coalition formation, unmodeled process in a sense. And so it looks like a really interesting mixture. You can break networks unilaterally, but you can't form them. You can break links, sorry, unilaterally, but you can't form them unilaterally. For me, looking at that paper, it has such an interesting mix of cooperative and non-cooperative elements. Did you see it that way or you saw it as really orthodox methodologically?
Matthew Jackson: No, no, certainly it was a departure. And I think in terms of the way we thought of it at the time, and part of the reason we used the term pairwise stability was when you're actually looking at this process of forming a friendship, for instance, there's a reciprocation that's necessary. And when you look at the way that non-cooperative game theory is structured, it doesn't have the right tools for working with coalitions somehow. It's very tricky to get a formulation of a game which captures the notion that two people can coordinate and communicate with each other and say, yeah, this is a mutually beneficial relationship, let's form it. So if you just have people call out which people they want to connect to, you get all kinds of extra equilibria. It's really a mess.
Rajiv Sethi: And it also feels a little artificial. It doesn't seem to capture the actual process of link formation.
Matthew Jackson: Exactly. And so what we were looking for was more just a stability notion. Which networks would we expect to endure? And it should be that nobody wants to get rid of relationships they have and nobody wants to add new ones. or two people that would mutually benefit from it don't want to add them. That simple perspective was something that had sort of existed in the matching literature. And so pairwise stability became a notion that sort of naturally fit to the network structure.
Rajiv Sethi: Yeah, that's really interesting. Maybe we can turn to some of the empirical applications. In recent years, you've been really looking at empirical applications. The original paper, I think, was largely theoretical, and much of your earlier work was. But you have been working with Raj Chetty of Harvard University on Facebook data about the role of network position or the nature of the people you're connected to in terms of your life prospects. Could we talk about that a bit?
Matthew Jackson: Sure. So I guess naturally, as one gets deeper into the literature and one spent more time modeling networks and so forth, you begin to really have a lot of conjectures about how they matter. And then specifically, I've been working on things like how do your opportunities matter in job contact networks? How do you learn from peers? How do social norms work on a network? So I built a lot of models. And the data— Facebook sort of offers a playground in terms of network data that's really unparalleled. Because you have connections between people and you can observe things about people. You can estimate their income, you can see whether they go to college. There's a whole series of things about people and you have a lot of friendship data as well. So it's a really wonderful playground with which to measure this. And so I started working with Raj Chetty when he was here at Stanford. We were just talking about, you know, he'd been working on opportunities and how your connections could matter. And so we started talking about whether we could really get into the network data. And Mike Bailey was a former student at Stanford. He was at Meta at the time, or then it was called Facebook. And so we started working with them. And then two former Stanford students who are now at NYU, Johanna Strobel and Teresa Kukler, were also working with Facebook data. So we all teamed up and started looking at the data, and it's just enormous. I mean, you have huge, we have 80% of the US population between the ages of 21 and 44.
Rajiv Sethi: And what do you know about them?
Matthew Jackson: So you know a lot about them. So you can see quite a bit, but what we were really interested in was trying to understand what affects a kid's trajectory. So if you look at, you know, a kid growing up, say, in the bottom half of the income distribution, in, say, rural Alabama, they're going to end up in the bottom half of the income distribution with high likelihood. If you look at a similar kid in rural Minnesota, that's not going to happen. Why? Well, that was part of the question was why, what's different about rural Minnesota from rural Alabama and so forth. And so what we started doing was looking at the networks. And so we said, "Okay, maybe it's social capital." So, you know, you can look at financial capital, you can look at human capital. So financial capital would be, you know, the wealth of the parents and so forth. That's more or less the same because we're looking at kids right in the bottom half of the income distribution, similar family background. You can look at human capital. You know, you can control for school districts and so forth. That's not doing a lot of the difference. But then you begin to look at the social capital. And that ends up being highly predictive. And so in the first papers, we sort of said, well, social capital is a term that covers lots of ground. It's been around. I mean, Glenn Loury will be here today at lunch. You know, he was one of the people talking about social capital. Bourdieu talked about social capital. Putnam's talked about social capital. Lots of people have talked about social capital, and it means lots of different things. And so what we did is we actually developed a set of different measures of social capital. So we worked with 9 sort of in the main paper. We had like 15 or so that we tested. But the idea was you can break it into different categories. You know, one form of social capital is your connections, bridging connections. Am I well connected in terms of opportunities, learning, information? Another is what people often call bonding capital. and this would be stuff going back to James Coleman, famous sociologist, how tightly knit is a community. So if a community has lots of links, it's really a cliquish community, everybody communicates with everybody, that can help enforce norms. So you end up with strong enforcement of norms, you can ostracize people, that can help. So that's another form of social capital. The third was more, I guess, Putnam, motivated, which is you look at local organizations. Are there volunteer organizations? Are people involved, active in a lot of different either NGOs or government agencies, etc.?
Rajiv Sethi: You're referring to Bowling Alone?
Matthew Jackson: Bowling Alone, exactly, exactly. So, you know, you could take these 3 different categories, and so then we developed measures for, you know, 3 measures in each of these categories and looked at them in the data. And to our surprise, one really, really stood out and the others were pretty much zeros or close to zero. And the one that really stood out was what we called economic connectedness. And so it's one of these bridging capital forms. And it was basically look at people in the bottom half of the income distribution in a given area and ask what percentage of their connections are to people above median income.
Rajiv Sethi: Is income the determining factor? Is education relevant at all?
Matthew Jackson: You can do it different ways. So we actually did different types of connectedness. You looked at age, you could look at— so it's somehow the economic well-being of the alters, of the people you're connecting to, seem to be the main predictor. And in follow-up work, that seems to be the real determining factor. So how well connected you are to people, look at it this way. If you look at a kid in high school, the biggest determination of whether they're going to go to college and what their future income is going to be is the average income of their friends' parents.
Rajiv Sethi: I'm glad you mentioned Glenn Loury's work, his dissertation actually. His most cited paper is actually this book chapter in a volume, I think it's called Women, Minorities, and Employment Discrimination. It was a 1977 paper. It's very, very interesting. Interesting because it introduces the concept of social capital into economics and also tries to think about these kinds of connections, especially if you have segregation along racial lines, which can explain persistence of group inequality. I've worked with Glen on that. But can we talk a little bit about the mechanism? So what could explain why the income of your affiliates' parents is so important in determining your own life prospects.
Matthew Jackson: Right. You know, so we've been doing follow-up work, and I think when you break it down, so look at, say, a cohort or a classroom, it matters much more who you connect with, who actually you befriend in those settings, than just the general set of people in the classroom. So it seems that It's not just that, you know, having more wealthy people around leads to a different classroom dynamic. It seems to be actually being friends with these people. And then, you know, at least hypothetically, there's a lot of pressures that all go in the same direction, which are, you know, if my friend has parents that are saying, "Are you studying?" or, you know, "Have you prepared for the SATs?" Are they thinking of going to college? That kind of information and set of aspirations is naturally transmitted.
Rajiv Sethi: Yeah, yeah, so aspirations are a key factor.
Matthew Jackson: Yeah, aspirations are a key factor. Just knowing, I mean basic knowledge about this stuff.
Rajiv Sethi: So I know that you and your co-authors have thought very carefully about causality, but people who are not familiar with that research are going to ask the question of whether or not those friendship formations themselves reveal unobservable characteristics about the people that would predict future success and the endogeneity of the friendship formation. And so you use random variation on some dimensions. So could you talk a little bit about that because the folks who don't know the work may think about, is this causal?
Matthew Jackson: Sure, sure. So there's a lot of reasons that in general working with network data is really tricky. And as you're mentioning, there's all kinds of reasons that we end up being friends. and those reasons could end up predicting our outcomes independently of the friendship.
Rajiv Sethi: Rather than the friendship itself.
Matthew Jackson: Exactly. And so in follow-up work that we will be releasing sometime soon, you can take, you know, you can look at high schools and look year to year, there's going to be variation, for instance, in the composition of a given class. And you can take advantage of the fact that the same kid in 1996 had a different opportunity in terms of connections of other friends than a kid in 1997. And so then you can take advantage of that variation and use instrumental variables and other techniques that economists are used to, to getting causal inference out. And it seems, you know, these correlations are so strong it's impossible to make them go away. So it's really, they really stand out extremely strongly in the data and they seem to be causal. It's pretty difficult to get them to disappear.
Rajiv Sethi: Yeah, that's really interesting. I wanted to cover that point because some listeners who are—
Matthew Jackson: Oh, sure. Yeah, yeah. No, it's very easy. I think, you know, more generally, one has to wonder what is it exactly about these friendships that ends up mattering? But I think a lot of things work in the same direction as we were talking about aspirations, but then future opportunities. You know, is this person going to help me get a job at some point? Do they help me out with medical and financial advice? I mean, there's all kinds of things that are lifelong that go through these friendships that matter.
Rajiv Sethi: Yeah, I think that it's very plausibly causal and you've dealt sort of statistically also with the causal question, but I think most people would recognize that those effects matter just from their introspection and daily life experience, I think. Yeah. Can we talk a bit about the Santa Fe Institute? So we're both on the external faculty, I think, and you've been involved with a project there. I think it's the Endow Project, is it? That involves people from other disciplines. So we started off this conversation talking about sociology. And here, I think you're collaborating with anthropologists. Is it Monique?
Matthew Jackson: Yes, yeah, indeed. So Santa Fe Institute, as we both know, it's a pretty amazing place in terms of the set of different disciplines it brings together. And naturally, this was another large-scale project that really is the brainchild of Sam Bowles, who we both know, who's at Santa Fe. And Sam is, you know, was trained as an economist, but is probably one of the most open-minded economists you'll ever meet. And he's been working with anthropologists over the years. And so he was interested in understanding inequality and how that relates to social relations. And so it's my interest and his very naturally dovetailed. And he had this idea of getting together Anthropologists go out and do fieldwork as, you know, often as part of their dissertation and then eventually their first book or something. They'll go out and spend time in some particular country and then some particular community gathering information about that community and so forth. So he was in touch with Monique Borgerhoff-Mulder, one of his collaborators, and a whole series of other people who were out in the field. And so the idea was, actually, let's collect the same style of data in a set of different societies. So now we have something like 50-odd that we're still getting data in, but it's, say, 51 societies around the world, rural, fairly isolated usually. There's some that are urban in southern India, but they're spread all over the world. They're into the Amazon basin, they're sub-Saharan Africa, they're in Mongolia, all over the world. And we collected exactly the same type of network data. Who do you share food with? Who do you exchange favors with? Who gives you advice? A whole series of different relationships. And then we track wealth of all these different households. And not surprisingly perhaps from the Facebook data, it turns out that when you look at the connections within a society, how those connections fall across different wealth levels ends up predicting how unequal the society is in terms of wealth distributions. So that's work we're still writing up at this point. But that involves sociologists or some— we have a couple sociologists on it, statisticians, a lot of anthropologists, a few economists. So it's a large-scale project and it's taken a number of years now. We've been at it for more than a decade. But I think it's going to offer unparalleled look at how the social structure relates to the wealth structure in these different communities.
Rajiv Sethi: Yeah. You know, it reminds me of an earlier project that also came out of SFI with which Joe Henrich, the anthropologist, was involved. And just to set the stage a bit, prior to the work of Henrich, Boyd, Bowles, Ernst Fehr, and others, With anthropologists, it was believed that certain experimental findings, for example, with regard to the ultimatum game, were universal. They had found very similar results in terms of proposer and responder behavior in the ultimatum game across countries. And the point made by Henrik and others was that, well, you're really looking at very similar populations. These are all educated. They're all subject to Western influences, they're really not that different. They're just living in different countries. When the anthropologists went out into the field, they found astonishing variation actually in the way in which people played games. Are you familiar with that?
Matthew Jackson: Yes, yeah, yeah, yeah, and especially as you point out, I guess Joe called these weird societies, these Western economic— so he's rich. We're basically looking— we think we're looking at diverse societies and we're looking at a a very specific set of them. And then when you start looking outside of that, you get very different results in terms of how people interact with each other, what their expectations are, and so forth.
Rajiv Sethi: So in a sense, that sort of approach to conducting research is a precursor to what you're doing with the NDF.
Matthew Jackson: Yes, definitely. So the idea of bringing together people across a whole series of different sites and then trying to do the same measurements, and Sam was part of that project.
Rajiv Sethi: He was. And Rob Boyd, the anthropologist. Rob Boyd, yes, exactly. Yeah, yeah, yeah. Quite influential, I think.
Matthew Jackson: It has, yeah. And I think it allowed us to actually recruit a lot of the anthropologists because they saw how influential that work was. And so it was easier to get people to be part of this project by saying, look, we're going to be doing this at even a larger scale with more data, different kinds of data, different questions. And it was easier to get people to be very excited to join.
Rajiv Sethi: Yeah, that's fantastic. I just would like to pivot a bit into your recent work with AI because it's also related to games like the Ultimatum Game. And I did not know about it until we started planning this conversation. And then you sent me a couple of papers that I found completely fascinating. And you have one that does a kind of Turing test for AI, but you also have AI agents playing games like the Ultimatum Game. The work of Henrik and the anthropologists in combination with economists looked at broader populations. And now you're broadening the population to AI agents. Is that what's going on?
Matthew Jackson: Yeah, I mean, I think of, you know, so I get bored easily. So there's lots of different things I get interested in. And AI, of course, is a topic that's so hot. But as a game theorist, you know, the natural thing that comes to mind is, We put people to tests all the time. As you're mentioning, this Joe Henrich and all papers, they were looking at how do people play across different societies. Well, we're trying to learn about the people from using these tools of games to see their behaviors. We thought, well, actually we can do that with AI as well. We have a whole set of games and different ways of evaluating people and trying to estimate how they're going to behave. Seeing what we think they're doing and why they're doing it. And AI is starting very rapidly to become part of our lives, increasingly in terms of giving us advice and steering, you know, our systems and platforms and advertisements and all kinds of things. We want to know how is it going to behave and why is it behaving the way it is? And can we see, you know, is there some way of evaluating its behavior systematically? And so we said, well, why don't we just do the kinds of things we're normally doing? And so this is work with Qiaozhu Mei, a computer scientist at Michigan, Yutong Zhi, one of his students there, and then Walter Huan, who's actually a PhD in biology but has been doing experimental economics for years and has a company.
Rajiv Sethi: These are published in the Proceedings of the National Academy of Sciences.
Matthew Jackson: Yeah, in PNAS. And so the first paper would say, let's just develop a methodology of using games and other kinds of revealed preference tools for studying AI and seeing, is it altruistic? Is it selfish? Is it risk-loving? So we can use the same kinds of tests that we normally have and just, for instance, put LLMs, various examples. We did ChatGPTs. We've done a bunch of different versions of AI.
Rajiv Sethi: So since this is so novel, can we talk a little bit about details? You provide a prompt that asks an AI agent how it would distribute a sum of money like in the dictator game. Is that what's going on?
Matthew Jackson: Exactly, exactly. So in the dictator game, you know, you give humans, you say, okay, you've got $100 to share with somebody else. How much do you keep for yourself and how much do you give to somebody else?
Rajiv Sethi: That's a standard dictator game.
Matthew Jackson: Standard dictator game. And that's some kind of measurement of altruism, fairness, and so forth. And you can ask, you know, an LLM, if you had $100 to give to another person, a person, how much would you keep for yourself? So it's a hypothetical, you know, question.
Rajiv Sethi: Yeah, it's fascinating. What do you find?
Matthew Jackson: And so actually we found— so a couple of comments first about it. So I think one thing to emphasize is the methodology can be used and it's very portable. So what we find about particular LLMs these days is somewhat ephemeral. So, you know, at that point we tried ChatGPT 3.5, 4.0. You can test newer ones. You can test Gemini, Lama, Claude, et cetera. What we found was in some games they were very typically human. So, you know, there's a fairly wide distribution of the way that humans act. So some humans are very selfish, they keep everything. Some humans give 50%, some give 30%. The chatbots were generally playing modal human behaviors. So they would, you know, give 50%. they were more altruistic on average. So when they tended to look non-human, they tended to look like they were more altruistic or splitting things more fairly, more 50/50 than humans tended to.
Rajiv Sethi: So I want to ask you a bit about the ultimatum game and then about another aspect of this project which tries to infer motivation. And so with regard to the ultimatum game, so slightly more complex than the dictator game, so the proposer will— propose a split, the responder can either accept or reject it. I'm just explaining this for people unfamiliar with it. Absolutely. And if the responder rejects, they both get 0, and if the responder accepts, then the split is implemented. First of all, what do you find with the AI agents in this scenario?
Matthew Jackson: Yeah, so actually, different AI agents will play slightly differently. They tend to be more generous in their responses than a typical human would be.
Rajiv Sethi: What do you mean by generous?
Matthew Jackson: Generous in the sense that, you know, some humans will say, look, unless it's 50/50, I'm going to veto it. You know, they might accept more of a wider range than a human would. But, you know, they're embodying human behaviors. It's one thing that's true about the— so we can talk, I think, transition a little bit. When you just give the instructions to one AI in one instance, it acts somewhat like one human subject would. And if we, you know, pull in like 100 different students, or like in Joe Hendricks' case, people from all over the world, you'll get very different behaviors. And so one thing you can do with the AI is you can, put in a system prompt before you elicit the behavior, which sets the tone for the whole session that you then go through. So instead of just saying, you're a helpful AI assistant, and now we're going to, you know, here's some instructions of a game, we could say, you know, think of yourself as a banker or think of yourself as a generous person. Think of yourself as— and so you can put in a prompt beforehand Can you say, think of yourself as an economist? As an economist, yeah. You could say, think of yourself as Rajiv Sethi or Matt Jackson, and it will embody what it knows about that role, and then it behaves very differently. And so it's not as if AI is sort of one entity. It depends on the context it thinks it's in, and that changes the behavior, and it changes it quite a bit. And so then we could study how much can you vary its behavior by varying those instructions.
Rajiv Sethi: And all the players in these games are AI agents. So in the dictator game, there's no other player. It's unilateral. But in ultimatum game, public goods, and so on, trust games, the responder will also be an AI agent.
Matthew Jackson: Yeah, you can put an AI agent there. You can also ask an AI agent to say, play as if you're playing from a poor person from a ghetto in Brazil, as opposed to, suppose you're playing from a rich person in New York City. and it will act differently. So you can sort of even tell it who it thinks it's playing with.
Rajiv Sethi: So what do you think we learn from these, these experiments?
Matthew Jackson: Yeah, so I guess two things. One is, in general, it's a way of evaluating AI, and I think we need to evaluate AI given how complex the code is. It's not— even if it's open source, the code is extremely tricky to understand. And then it's trained on large amounts of data. So one is that we can use these tools for just saying, okay, let's put it through a bunch of different tests and see how it's behaving and try and in general understand what motivates it or what it's as if it is motivated by. And the second is that we can also use some of this, we, you know, by varying these prompts, we can learn a little bit about what different human circumstances might be and what might be motivating the humans. So if we say, you know, act as if you're risk-averse, it changes its behavior in different games, or act as if you're fair-minded, it changes its behavior in different games. And so it's a way of testing a lot of things that economists have theories about, fairness, risk aversion, and so forth, putting those into the prompts, we can see how it varies its behavior and that gives us a way of trying to understand what might be motivating humans and what might be changing human behavior.
Rajiv Sethi: I want to ask you something and tell me if I understood incorrectly the work that you sent me. So some of it deals with making inferences about human motivation. Is it something along the following lines that you can presented with results of human behavior, let's say in the ultimatum game, and it tries to infer what might be driving that behavior? Is that something?
Matthew Jackson: Yeah, so the way we actually do it is, so suppose we've got a range of different behaviors.
Rajiv Sethi: From human experiments.
Matthew Jackson: From human experiments. So let's take, for instance, the Joe Henrich and say, okay, this is the way people acted in sub-Saharan Africa. This is the way they acted in North America. What prompt do we have to give it to get it to act like the people in North America? What prompt do we have to give it to act like the people in sub-Saharan Africa? And for instance, if you say something like, you're more civic-minded, you're fair, you'll get more behavior that looks like sub-Saharan Africa. If you say you're more selfish, you're more self-interested, you'll get things that look more like, say, North America. So you end up with different kinds of things you have to feed into it to get it to match different behaviors. Then the question is, can we interpret those prompts that we're giving it?
Rajiv Sethi: So it's not as if you are just presenting it with the data and asking it in an open-ended way to explain motivations, but one could in principle.
Matthew Jackson: You could, yes. I think that's to some extent sort of like asking people after the fact why they acted the way they did.
Rajiv Sethi: Yes, except that you're asking the AI agent rather than the people themselves. The reason I find this kind of interesting is because in the ultimatum game, there's two very different possibilities for proposer motivation. So one could be that, look, I don't want to present an unfair allocation to the responder because I just don't like to be unfair. The other is that some responders will have put such a priority on fairness that they'll reject even slightly unfair offer, so let me not take the risk. Those are two very different motivations for the same kind of behavior. This reminds me actually of an incident, I don't know if you've heard of this, where the early ultimatum game experiments, the Guth, Schmittberger, and Schwartz experiments were done I think with 6 or 8 or 10 Deutschmarks. This was the 1980s and it was a very small amount even in those days. People made the argument that, "Oh, you scale it up, make it some serious money and you'll get behavior according to the prediction of subgame perfect equilibrium based on material self-interest," right? The classical prediction which did not take into account preferences for altruism, reciprocity, and so on. Alvin was asked, "Well, how would you play if it was a million dollars as a proposer?" And he said, "Oh, obviously, I'll offer half a million." And people said, "Why? I mean, you're an economist.
Matthew Jackson: You're a game theorist.
Rajiv Sethi: Why are you doing this?" And he said, "You never know what kind of crazy person you're encountering, right?" And basically, half a million is a lot of money. Why risk it even— And so it was a very different motivation than fairness or reciprocity. It was just a sort of diminishing marginal utility type explanation. And I'm just wondering whether, given any dataset, the potential exists for AI to make inferences about motivation independently of variation of the prompts. I'm just curious about that.
Matthew Jackson: Yeah, I don't know the answer to that. So it's hard. I think even when you ask humans, like ex post, suppose somebody said I gave 50% they have a temptation to say, I gave it because I was generous, because that makes them— even if it was risk aversion, even if it was because they, they were just worried that it was going to be turned down, right? And so it's tricky after the fact to interpret these things. And AI, I think the, the advantage of AI in some respects is the sheer volume of human experiences that it's basically ingested in its training data is enormous. And so it can it can process all of that and begin to understand what might be motivating humans. And so I think querying it in different ways, you know, we've been working it from the prompts itself by saying, what do we have to give it in terms of motivation to get these different behaviors to match? But yeah, you could do it the other way around where you sort of say, you know, what's going on? One thing that was fascinating that we found in some of our experiments was If you ask it, say we give it a prompt and then we ask it to explain its behavior afterwards, but we tell it that we're going to ask it to explain its behavior upfront, it'll act differently than if you don't ask it to explain its behavior, right? So you just sort of say, you've got to do something and give us a motivation for why you're doing it. It'll give a different answer than if you just say, you know, play.
Rajiv Sethi: If you tell it that it has to provide a motivation before you ask it to conduct the task.
Matthew Jackson: Exactly. So you say, you know, give us your play and tell us your motivation. It'll give a different answer. Humans are trained on data. We're trained on all kinds of messy datasets. And it's just that these are sort of humans on steroids in the sense that they're able to get a much wider set of data. But we're trained in all kinds of imperfect ways. We're seeing misinformation, disinformation, processing it, trying to make sense of it.
Rajiv Sethi: But you know, going back to your networks, you can think of LLM as a network of individuals, as embodying data from a network of individuals, including anyone who's ever lived and left anything behind, you know, going back millennia. And so you're really thinking of somebody with centrality, if you like, that's orders of magnitude more than any human centrality. You know, that's one way to think about an LLM. And that's a sort of a network connection. Right.
Matthew Jackson: I mean, there's some, I think probably both neuro and other evidence that even humans are somewhat, we think of ourselves as one person, but we're actually a complex net of processes that are interacting.
Rajiv Sethi: Well, neural networks are used to simulate the brain, right?
Matthew Jackson: Exactly. And to some extent, LLMs are just larger, more complex versions of that in some dimensions and more data, but then simpler in other dimensions. So actually, the human brain still seems to be more efficient than the LLMs are in the transformer technology in terms of numbers of nodes and the wattage that we can get by. it's quickly going to be changing on that dimension as well. And, you know, similarly, humans are trained just by getting feedback from— we form a belief, we take some action, and then we get some feedback for it, and then we revise somehow. We don't even know exactly how we're revising. But LLMs can be trained in the same way, right?
Rajiv Sethi: Yeah, I think it's worth taking a minute to I wonder if I could ask you the following question, which is that, because it links to networks, you know, and is it a useful way to think about a large language model or, you know, AI as being an agent with a massive number of network connections dating back to anyone who's ever existed and has left behind anything on which the model could be trained and therefore has a sort of a very privileged network position more than any human in history ever has had. Is that a useful way to think about what an AI is?
Matthew Jackson: Yeah, actually it is. I think it's— so definitely the way Transformers work, I think that is right in the sense that the way in which it can take data can produce a network structure among the different aspects of all the different data it's processing.
Rajiv Sethi: Yeah, because it's sort of drawing on— if you think about everything that's gone into the training as arising from a friendship network, you know, where the AI is sort of linked to what, you know, Thomas Aquinas wrote or Aristotle or, you know, whatever, is all in there.
Matthew Jackson: And it's putting it in different contexts. It's putting it— it's weighting it and digesting it in a complex way.
Rajiv Sethi: Which is kind of what we do in a way that's not transparent to ourselves. So Matt, I'll come back to AI a little bit later. As my last question, I have something I wanted to ask you, but I'd just like to ask you a little bit about the process of writing books. You have two books that I know of. One is an academic book that's really requires some technical background to appreciate. Then I think that one is with Princeton University Press. Then the other one is what's called a trade book where the audience is broader and you're trying to bring this research to non-specialists. Could you just talk a little bit about the process of writing books in general, the differences between these type of books, and whether or not economists should be writing more books? We tend not to now.
Matthew Jackson: Yeah, I mean, I think so. First of all, you're absolutely right that the difference in these types of books is enormous, both in terms of who it's reaching and what it takes to write the books. And they both present very different challenges. And I think they both were very useful for me. So the first book, the one for researchers, on social and economic networks was one that I mainly wrote when I was up here at CASBS in 2005 and '06. So that was one of my main things I did here. And that one was sort of trying to synthesize, especially in networks, there's literature in physics and sociology and economics. And I was trying to bring together all of the different tools and all of the different models and the different approaches and sort of with an economist view of application. And sort of just pieced together what we knew. And it was very useful for me in terms of synthesizing. So it's a very useful project in order to force you to try and make sense of whole literatures and whole approaches that people have used and what are the powerful things.
Rajiv Sethi: So it wasn't just a service to the profession in the sense that graduate students could learn a huge amount from a single book. It also helped help you to clarify your thinking.
Matthew Jackson: Yes. Yeah, yeah. I mean, it helped me reorganize my thoughts and a lot of my further thinking about things like centrality measures and why there are different measures and typologies of them. And there were a whole series of papers that I wrote later that came out of that and sort of rethinking the literature and thinking where there are holes and what we've done well and what we were missing. So I think it's very useful there, but it was— I was intending as a toolbox for graduate students, basically because there wasn't something really easily available for them at that time. And so that was the purpose of it.
Rajiv Sethi: And what about the other book?
Matthew Jackson: The other book actually came out of the first one. My wife was helping me proofread it, and she would read through these chapters and say, wow, you know, the first 2 pages are really interesting where you do the introduction, and then you get into all this math. And so she said, why don't you write a book that's just expanding on these first 2 pages of each chapter. And so that was the idea, was sort of bringing what we've learned in network science, economics in particular, to the general public. But that presents its own challenges. So when you start working with editors and you talk to them about— I had something I hadn't realized. As an academic, we write very elaborate sentences, and the sentences usually start if, you know, a certain set of conditions are met, then we can make this conclusion. And if you write that sort of qualified conditional way, it's not a great way to communicate things to a general public. And it's very turgid, it's complicated, it's laborious, and it's very precise, but it's not an easy way to communicate ideas to a general public. And so then you have to find examples that make clear all the nuances and why you need all these ifs and so forth. But you have to unpack it in a different way. You want to sort of make statements that are clear and understandable and easy to unpack. And so it takes a lot more time in some sense to find those perfect examples and illustrations of concepts and trying to explain things without all the extra baggage that we usually have in the academic world.
Rajiv Sethi: Yeah, I recognize these challenges because I'm trying to work on a book that that is on that sort of trade academic interface. But that's very helpful. Thanks so much. Can I ask you a little bit about the contrast between CASBS and the Santa Fe Institute? They're both located in absolutely beautiful places. But CASBS has the year-long fellowship similar to the Institute for Advanced Study at Princeton or the Radcliffe Institute. And Santa Fe has short visits. My experience there is, you know, visits can be as short as a week or less even or much longer. Offices tend to be shared. Doors tend to be open. There's an endowed tea that, you know, everybody gathers around at 3 PM. And it's just a very, very different way of sort of cross-fertilization across people and interaction. I just wanted to ask you what you think about these different approaches to getting sort of serendipitous interactions that can generate interesting new research. Which model do you think really is most fertile?
Matthew Jackson: It's actually interesting because I think they've both been successful models and they come out of different traditions and I think you can trace part of the structure to those traditions in the sense that CASBS was more coming out of a sociology, social science perspective, where people were often writing books. And originally, people would come up here and, you know, as I did and you're doing, you know, you spend time writing a book and then interacting with other people, and it helps broaden your perspective, and you have interactive seminars and so forth. But you spend a lot of time deeply thinking about problems. You might be bringing in a working group where they spend a whole year on a particular topic. Whereas you're pointing out, Santa Fe Institute came more from a physics complex systems background originally, but then also mixing in people like Canero and others who—
Rajiv Sethi: Yeah, a lot of the proximity to Los Alamos played a role, I think.
Matthew Jackson: Yeah, yeah. And it was much more built off of, okay, we have topics and workshops, we bring in key postdocs who are working on stuff. The postdocs will be some glue, but we want sort of fertile, very creative, unusual thinking kinds of projects and a little more high-risk kind of stuff. So it's a different model, and I think they've both been very successful in different ways.
Rajiv Sethi: Yeah, I think they're both absolutely fantastic places. On SFI, I just want to mention something in case there are graduate students listening. So when I was a graduate student, I would just browse around in the basement of the Strand Bookstore in New York. You know, they had relatively low-priced books, review copies and so on. And I came across this book called The Economy as an Evolving Complex System. And it's edited by Ken Arrow along with two physicists, Anderson and Pines. I look at it from time to time. It's completely astonishing. It just changed the way I look at the world and it is so different from the economic methodology which is very focused on characterizing equilibria. This book had contributions by Stuart Kauffman, John Holland. It had an incredible paper by Mario Henrique Simonsen who was once a Brazilian, I think, finance minister maybe. I hope I got that right. That's where I first saw the Half the Average Game. This is well before Rosemarie Nagle's experiments were published in the American Economic Review. That book opened my eyes in a way that I think can be useful even for current graduate students. I mean, neural networks, optimization on rugged landscapes, these are things that economists should pay more attention to and after all these years, I feel that it hasn't had as much impact as it probably should. Are you familiar with this stuff at all?
Matthew Jackson: Sure, sure. And I think, I guess, a couple of observations. So one is that when you think about economics graduate student training and so forth, I think one powerful aspect of economics is that it has had a pretty tight, narrow paradigm. So, you know, the kinds of— if you go to different PhD programs around the world, a lot of it would be very very similar in terms of the first-year course sequence.
Rajiv Sethi: That's it.
Matthew Jackson: Right? So you'd have, you know, Moskowitz and Green, where it might even be the same textbook that they're using across these things, which teaches them, here's, you know, preference theory, and then we have equilibria, and we have implications of those in a bunch of different settings and so forth. And that's very powerful in uniting economists in having a common methodology and a common base from which to depart. But it also means that there's some narrowing in terms of the things that people tend to be exposed to, given how much you have to get into them in a couple of years. And a place like SFI opens your eyes to say, wow, there's a whole set of different perspectives and different toolboxes out there that are very useful in building different types of insights about complex systems and the way that they work. And similarly here, I think this was an eye-opening experience when I spent the year at CASBS, and sort of thinking back to say, why do we approach problems the way we do in economics? Why are we always starting from some kind of maximization problem or some particular paradigm? Maybe there's other ways to look at the world and understand it.
Rajiv Sethi: And I think in graduate school may be the time in which to expose oneself to it. Of course, then it's risky to then go entirely off the reservation, so to speak, and people need to convince their peers that they're worth employing and so on. But my own experience with the constraints of standard methodology really comes about through my work with Muhammad Yildiz and trying to move away from the common prior assumption and all the paradoxes it creates like no trade theorems and the absence of public disagreement and so on and so forth. We just pushed the boundaries slightly just working. And heterogeneous prior models are now pretty standard now. But when we first started this, actually, he was my next door— he had the office next door to mine at the Institute for Advanced Study during a fellowship year. And we just got talking actually about different reactions to the O.J. Simpson trial. And he convinced me that what was going on could be usefully understood through a heterogeneous prior framework. But that's a sort of a marginal or minimal departure from the standard framework, but the stuff in this Arrow, Anderson, Pines volume, this is 1987, they're talking about rugged landscapes and neural networks way back then. That was kind of mind-blowing to me then and remains so actually now.
Matthew Jackson: Yes, yeah, yeah, yeah. I think Ken Arrow was obviously a sort of unique individual.
Rajiv Sethi: He really was.
Matthew Jackson: Really seeing the power of different approaches and so forth.
Rajiv Sethi: He really was. And we talked about Glenn Loury a while back. And he, I think yesterday or the day before, gave the Arrow Lecture here at Stanford. Yes.
Matthew Jackson: Yeah, yeah, yeah, yeah, yeah. It all comes full circle.
Rajiv Sethi: He also gave the Arrow Lecture at Columbia. There's a Columbia Arrow Lecture because I think Arrow got his—
Matthew Jackson: We have two lectures here on campus, Arrow Lectures. Yes.
Rajiv Sethi: Yeah. Columbia tries to claim Arrow or some part of Arrow.
Matthew Jackson: Harvard does as well.
Rajiv Sethi: Yeah, yeah. Harvard does as well. So I have one last question, which is a kind of a big question. It's not really directly related to the research we've been discussing, but it is linked to AI. It's rapidly transforming the way in which we do research. And I just wanted to get your thoughts on what the world might look like maybe 10 years from now. First and foremost, maybe our world, our professional world as scholars, social scientists, but then more broadly, the world at large in terms of you know, the economy, what we're producing, how we're producing it, and so on. You know, if you have any thoughts on that, I think it's a question on a lot of people's minds.
Matthew Jackson: Yeah, I mean, I think it's probably one of the more important questions in the coming decades. One observation is that when you look historically, and there's two things that are different about this technological innovation than previous ones. So previous ones, were somewhat narrower in the scope of application. So when you think about mechanization, industrialization, you think about bringing in the plow and replacing horses. There's a whole series of—
Rajiv Sethi: Electrification.
Matthew Jackson: Yeah. I mean, a lot of these played out over centuries, decades. It took a long time to play out, and they were fairly focused in the parts of an economy that they were interacting with. AI seems to be something that has a much broader reach. So it can be, you know, it can be redoing the way that legal contracts are written, as well as the way that economists are gathering data and analyzing it, and the way that platforms are advertising. So it's sort of widely—
Rajiv Sethi: And more rapid.
Matthew Jackson: And more rapid, yeah. So we're going through things, and, you know, even You know, the sort of computerization and web stuff took decades. This seems to be accelerating. So every time you blink, there's a new innovation coming out. And whether we're going to get AGI or, you know, these kinds of things, it could be happening relatively soon. So I think it's faster, it's broader, and it means it's going to have an enormous impact on our lives. And for good and bad, in the sense that I think each one of these technological revolutions you look at, had enormous benefits in overall productivity, GDP, but they were also very disruptive. So you had recessions, depressions, you had widespread unemployment in certain sectors of the economy. And so we'll be facing some of that.
Rajiv Sethi: One thing we've been seeing already, actually last year, was a disconnect between GDP growth and employment growth. So output has been growing. Faster than employment, but whether that's temporary or whether that's a sort of long-term structural change—
Matthew Jackson: Yeah, it could be an early indicator of what's to come.
Rajiv Sethi: Yeah. Matt, I just want to thank you again for joining us. It's always a pleasure to talk to you.
Matthew Jackson: Oh, it's great talking with you, Rajiv. Thanks so much.
Narrator: That was Matt Jackson in conversation with Rajiv Sethi. As always, you can follow us online or in your podcast app of choice. And if you're interested in learning more about the center's people, projects, and rich history, you can visit our website at casbs.stanford.edu. Until next time, from everyone at CASBS and the Human Centered team, thanks for listening.