Sean Carroll's Mindscape: Science, Society, Philosophy, Culture, Arts, and Ideas - 330 | Petter Törnberg on the Dynamics of (Mis)Information

Episode Date: September 29, 2025

A characteristic of complex systems is that individual components combine to exhibit large-scale emergent behavior even when the components were not specifically designed for any particular purpose within the collective. Sometimes those individual components are us -- people interacting within societies or online communities. Studying the dynamics of such interactions is interesting both to better understand what is happening, and hopefully to designing better communities. I talk with Petter Törnberg about flows of information, how polarization develops, and how artificial agents can help steer things in better directions. Blog post with transcript: https://www.preposterousuniverse.com/podcast/2025/09/29/330-petter-tornberg-on-the-dynamics-of-misinformation/ Support Mindscape on Patreon. Petter Törnberg received a Ph.D. in complex systems from Chalmers University of Technology. He is now an Assistant Professor at the Institute for Language, Logic and Computation at the University of Amsterdam, Associate Professor in Complex Systems at Chalmers University of Technology, NWO VENI laurate, and senior researcher at the University of Neuchâtel. Web site Univ. Amsterdam web page Google Scholar publications Amazon author page Bluesky

With the new My Lowe's Pro Rewards American Express card, earn three points per dollar spent on Lowe's qualifying purchases for the first six months after card account opening. Plus, save 5% every day on eligible purchases at Lowe's and earn points on eligible purchases everywhere else MX is accepted. Card members get more at Lowe's. Subject to credit approval. 5% can't be combined with any other offer.

Exclusions and restrictions apply. Points subject to loyalty terms at Lose.com slash terms and credit reward terms at synchrony.com slash MLPRA dash terms. Visit Lose.com slash business credit. Indeed sponsor jobs gets you quality candidates when you need them most. Spend less time searching and more time actually interviewing candidates who check all your boxes. Less stress, less time, more results. When you need the right person to cut through the chaos, this is a job for Indeed sponsored jobs.

And listeners of this show will get a $75 sponsor job credit to help get your job the premium status it deserves at Indeed.com slash podcast. Terms and conditions apply. Need to hire? This is a job for indeed sponsored jobs. Hello, everyone, and welcome to the Mindscape podcast. I'm your host, Sean Carroll. There's an idea in social science circles called physics envy. Economics especially is susceptible to this idea.

It's not supposed to be a good thing. You're not actually supposed to be envious of physics, but social science is hard. People are messy. There's a lot of variables going on. Physics is able to make enormous progress by simply. simplifying things a great deal. In part, that's because the fundamental ingredients that we study, even though it's like quantum mechanics and cosmology and relativity and other things that sound way out, but there aren't a lot of moving parts. The basic things that we're looking at are sufficiently

simple. You can describe them using relatively few variables, and you can isolate all the interesting things that are going on in these systems with small numbers of variables. As a result of this, you can make tremendous progress. You can prove theorems, you can do experiments, the test your theories to many, many decimal places. It's a lot of fun. Of course people would be envious of this. But it's a disease, or at least something to be avoided, to therefore try to make your social scientific research too much like physics. When you do social science, you should admit that there are complications there that cannot be abstracted away in the same way that we abstract away air resistance or friction when we're doing physics. Nevertheless, I'm

sure that everyone who listens to Minescape on a regular basis knows, I do think that there are contexts in which physics-like reasoning can be helpful or even interesting in the social science. There can be contexts in which physics type of reasoning and concepts borrowed from physics can be really useful, very interesting in the social scientific contexts, ideas like equilibrium, ideas of emergence in general, ideas of what is collective behavior like when, it arises from the sort of mindless, non-directed interaction of many small things. These are things that physicists think about all the time and are very, very relevant to the social sciences. So today's guest is Petter Turnberg, who is a professor of computational social science.

I promise I didn't know this, but he admits on the podcast that he actually has a physics background, so this makes some sense. But he uses models, agent-based models that we've talked about recently with Don Farmer and other. as ways to study the behavior of social systems. Can you make a little model where the individual pieces are either simple agents that always act in some way, or maybe there's a little bit of stochasticity in there, or maybe they're even very complicated.

We'll talk about an example where better used LLM's, large language models, to model human interactions in social media landscapes. And then you can ask, what is the robust behavior? Do you get things that we observe in the real world? Do you get polarization? Do you get sort of an accumulation of influence in certain people rather than having it be completely uniform?

Is this good or bad? If you intervene in certain ways in a social medium, can you make things better? Is it all the algorithms fault? Or is it just the preferences of the individual actors? All of these kinds of questions can be addressed with this sort of slightly physics-y attitude, but put into a social science context. I don't want to give away too much about Petter's results, but they're not great.

They're not very encouraging for those of us who want social media to work. There are certain natural dynamics that seem to come up that drive things in a bad direction, that drive polarization and echo chambers and things like that. It doesn't matter whether we like it or not, if that's what happens, knowing how it happens and that it happens under certain circumstances, hopefully will be helpful in making social media

or media or information ecosystems in general more functioning for the purposes of democracy, but also for the purposes of just learning, fun, new things, and having a good time and being good human beings, in connecting with our peeps in various different ways and building connections that otherwise wouldn't have been possible. So we want to keep the good aspects of these wonderful technologies without being subject to the bad aspects.

And I think this kind of study helps us figure out how to do this. So let's go. Fetter Torenberg, welcome to the Mindscape Podcast. Thank you for having me on. It's a real treat as a longtime listener and fan of the podcast. It's really great. Okay, very good. Well, then you know how it's going to go.

Let's just start setting some stage a little bit. In one of your books, you have this provocative line, Power has an epistemology. Do you remember that one? That's from seeing like a platform. So what does that mean to those people in the audience who might not use words like epistemology every day? Yeah. So it's a good starting point, I think.

So a lot of my research is kind of informed by this understanding of society being a complex system. So I kind of like I actually have kind of physics background myself, but like long time ago. So don't quiz me on that. but and come very much from the kind of complex systems perspective. And in my PhD, I was kind of focusing on taking that perspective to try to see how we can understand society using the methods of complexity science and using computational methods. And so this this book is basically centered around this idea. Okay, we're seeing this notions of complexity becoming used throughout the social sciences. But also in kind of urban planning.

you know and also in kind of when we talk about digital platforms and how they're shaping society and so we try to kind of understand what does that actually mean and so i mean in the social sciences the way the complexity tends to be understood is you know like this bottom up system so often you have like you know there's complicated systems like a car or a spaceship that you can kind of take apart you can decompose them you have like an engine and it's quite easy to figure out how they fit together. You can kind of understand the system by taking it apart. But then on the other hand, on the other side of this line, you have complex systems. So like ant colonies or flux of birds or whatever. And these systems, if you take them apart, you know, like you take out an individual ant from an ant colony, you can observe its behavior as much as you want, but it won't tell you very much about how an ant colony functions because the kind of intelligence of the ant colony emerges through the interaction of a lot of ants, right? And so the book kind of stems from the observation

that we've shifted to more and more talking about society through this complex lens and that is kind of intertwined with this notion of new forms of democracy because obviously a lot of those ideas come from kind of from, you know, from your field, from physics and from computer science. But as they're entering, they're changing quite a lot

as they're entering into the social world. And they also begin to have kind of political implications. And so those are the kind of implications that we follow. And ultimately, in the social sciences, this becomes a question of kind of epistemology. So the question of like, how do we envision what society is? And we went from, you know, in the 60s and in what we use, social scientists, referred to as Fordism or industrial modernity, where we tended to see society as a machine as a complicated system. So basically, that way of seeing and way in understanding society stem,

we argue from a society built on large industry, large mass production, and it led to kind of a mass society. And it led to the kind of ambition of like we can design, we can plan society, we can build it as if we're building a machine. So this is Fordism as in Henry Ford and his assembly lines. Exactly. So that's that image of the factory that Ford produced because Ford, he didn't only produce a kind of way of producing, but also a way of consuming. And social scientists have looked at how that kind of idea of the industry, how it kind of leaked in society, leaked into society and shaped how schools functions, how companies function, you know, like that it became a kind of way of organizing society, broadly speaking. And so what we're kind of observing in this book is

the fact that we've moved from that kind of machine epistemology into an era that's defined by kind of complex systems where we're talking about society as kind of swarms or as self-organizing. We're using these different metaphors that stem from, they're much more organic, you know. They're not the kind of machinery, you know. And we've also kind of in some ways abandoned the ambition of having the state to design society to produce certain outcomes. So this kind of utopia, this hope of improving the world, we've kind of given up on that. And instead, and there is this idea that, you know, now that it's implying that the outcomes of systems, like the bottom up outcomes of systems somehow would be inherently better. That's the kind of underlying assumption of this, what you might call ideology.

And so that's kind of what we're interrogating and questioning because we would argue that this isn't just something that's natural. There are still forms of power that are shaping society. They're just much less visible. And they're operating through by shaping kind of how we interact by things like algorithms. And so they often become kind of difficult to see, but they can still have very large structural outcomes on society. So for instance, it's by kind of fiddling with the rules of interaction, which is basically what platforms are doing. And that can have really important large-scale outcomes. and there is still power there.

It's just power that has a new epistemology. So that's very, very helpful. Thank you. I mean, I read the beginning of your book, but I didn't get through the whole thing. So this is why we have the podcast, so I can just ask you questions. So in other words, let me try to rephrase it and see if I'm understanding. You know, we might have had this dream of planning an organization where it was top down and either Henry Ford, who I guess was very capitalist at heart, but still, you know,

he was trying to organize his factories or, you know, central planning from a government. And there are various arguments you can make that simply letting things come to an equilibrium is more efficient, whether it's a sort of thermodynamic equilibrium or an economic free market equilibrium or whatever. And you're making the point that maybe so, maybe that's sort of a better way of calculating some optimum, but it still carries with it. its structures of domination and power. For sure. I mean, there is this, you know, even emergent outcomes,

even when we arrive at an equilibrium, it's still like those outcomes are still defined by the conditions that we came in with, right? And so, I mean, it's like these arguments that you sometimes hear from certain parts of academia where it's, they build some model of the economy and they find that certain people become very poor or like in large majority become very poor,

and certain people become very rich. And they say, well, it's an inevitable outcome of this system. We shouldn't try to change it. But that's just not a good argument. It doesn't follow, right. Yeah, it doesn't follow because you could have produced other rules that would produce other outcomes. And if your outcomes that are produced by your system

are problematic and are harming a lot of people, then maybe you should reconsider the rules that you're operating under. And also with the notion that like this idea that the market would somehow be natural and not, you know, like that it would not be shaped by state power has been, you know, questioned for the better part of 100 years that actually the market is very much a construction of state power. I remember the epiphany I had when I was thinking about different probability distributions. This is purely a math statement, but still it's, you know, important. Like we have uniform probability distributions, power laws, bell curves, whatever. And you might ask, well, which, one maximizes the entropy. And the answer is they all do subject to different constraints. Like it's it's all about the constraints you put on the system macroscopically.

Nothing is there really inevitably. Yeah. No, and I think this, I mean, it's a really, this is very much at the core of what I am interested in to certain degree. And I think it's also very much at the core of many of my papers and much of my research on this, which is the fact that as we've moved from a kind of a spatial society, you know, we moved from a society that we could, you know, if we were, as a physicist, one might think of it as a kind of lattice, you know. Our nearest neighbors. Yeah, exactly. And then we moved to a digital society, which is characterized by network structures.

And those produce, you know, like they are associated in different types of distributions. As social scientists, we, you know, everything is a bell curve, right? Yeah. But, you know, as a computational social scientist, everything is power law. Everything is power laws, right. And that's not just like a question of how we study these systems and what assumptions we need to make. It's also, you know, that certain people, when we have digital network structures, become very powerful.

And most people are, you know, powerless. And they don't get attention. They don't get resources that are important. And, you know, that is profoundly problematic for society. and they are attributes of these networks or the structures. We shop for everything at home now. Why can't we shop for blinds at home too? That's why I love three-day blinds.

Instead of wandering around a store, three-day blinds sends a professional design consultant straight to you. They bring all the samples, help you see what actually works in your space, take expert measurements, and even handle the installation. We've all bought something that looked great in a store and looked terrible once we got home. With three-day blinds, you see everything in your actual light, with your actual furniture, before you commit.

Plus, whether your home or away, you can control your window treatments with the three-day blinds app. So for a free, no-charge, no-obligation consultation, head to three-dayblinds.com slash pod-50 for our buy-one, get-one, 50-per-off deal. That's the number three, DAYblinds.com slash pod 50. I mean, maybe a good inroad here, and I think you just gave us a good segue to it, but there's this classic work on self-organization in society by Thomas Schelling and his segregation model. So you have an updated version of that, but why don't you tell us what Schelling's version was? Sure. I think it was published in 1969 already.

And it's still kind of pointed to, I mean, I use it all the time when I'm teaching, because I think it's still somehow the best model. It's not just the first. Also, we peaked as computational social scientists. And so basically, Shelling was kind of, he was, I think the backstories that he was like walking through his cafeteria in the university and was looking around and he saw like, it's weird that, you know, all the geographers are sitting in one table and sociologists are in another table.

And then he was like, yeah, and the same thing with the city. And he was trying to figure out like, why is segregation? such a common outcome across systems. And so what he did, he actually took a checkers board. So he wasn't actually, you know, he wasn't even a computer simulation back then. And then he had coins on the checker boards. And he was, okay, so let's imagine that the system is a lattice.

And agents are randomly distributed on this checkers board. And each agent, they follow a simple rule. So if more than like a very high number of their neighbors, they're like more neighbors are of a different type than themselves, because they're two types of coins, then they move to a random available space on the lattice. And so, and you can have the rule of something like that. They're completely satisfied if, you know, 70% are of a different type than themselves.

But if it's like 90%, they're like, yeah, okay, like, you know, I'm tolerant, but like, come on. Right. So they want some neighbors to be, they want some neighbors to be like themselves, but they don't insist that most of them be like themselves necessarily. Exactly, exactly. And then, so they moved to a random space. And so what happens? You would kind of expect that if they're happy with, you know, like 80% being of a different type, you would expect the system to settle on maybe something like 80% or, you know, like, pretty much 50-50. You wouldn't expect very high.

levels of segregation to emerge. But what actually happens is that you get almost complete segregation, even with very kind of high thresholds. And why is that? Well, basically, you get this kind of cascade effects. So one person is leaving the neighborhood. The neighborhood is left more segregated. His like neighbors also move and you get this kind of cascade. And so basically as a, what the model is telling us is that the integrated state is just very unstable. And so the system tends to tip over, like, any neighborhood tends to tip over to one color or the other. And so to me, looking, I mean, it's always been my favorite model. And to me, that seemed to tell us something also about the digital world.

So I was interested in, like, can we generalize this to other type of interaction structures that are non-spatial? And so I look at kind of forums and different platforms because basically, When looking at platforms, when looking at social media, there's been like for the last 20 years a debate on, because we often see echo chambers, right? We often see spaces being very homogenous. And then there's been a longstanding debate about whether that is driven by the algorithms, like the filtering algorithms. So there's this notion of Pariser's notion of from his 2011 book, The Filter Bubble. So this idea that the algorithms create this cocoon. They only show us content that the same like that we already agree to.

And then especially in recent years, there's been more and more argument that that's not true, that in fact, we want to be segregated. We don't want to, you know, we don't want to be exposed to other ideas. We don't want to encounter someone that disagree with us. And so that's been the kind of the debate between those two positions. And to me, it seemed like that might, you know, maybe it's neither. So basically what I do in this paper, it's a very simple, in short paper, I implement the Schelling model,

but I move it online to certain degree. So I look instead of having a lattice, I look at, you know, we have different groups so that can represent kind of subredits, it can represent websites. And the agents are randomly located. So it's again very similar to the Shilling model. So it's just as simple as possible.

So the agents are randomly located to these groups. And then in each round, they interact with some, some random people in their group. And then it's the same rule as in the shelling model. So they're happy with, you know, as long as at least some of their interactors are of their same type. And if there is no one of their same type, or maybe a very few percentage,

they move to a random other group. And I mean, with this background, I can kind of guess what the conclusion is, but what I find is that actually the shelling, chelling segregation effect is even stronger in this kind of communities, that they're even more prone to segregate. And of course, that has kind of interesting implications for this debate, because it's not necessarily either that this is driven by filter bubbles, that is driven by algorithms, nor is it in the interest of anyone. It's just something that follows from having

social interaction and structured in these ways. And there's also some kind of counterintuitive results from this. So for instance, actually having filtering algorithms, having a filter bubble, it actually reduces segregation. Because if you have a filtering algorithm that always shows you someone who agrees with you, like, or whatever you're shown, you know, some messages, you always get someone who agrees with you put into those messages. You will become less prone to move. To moving. Okay. And so that will, like for the system level, it will reduce the amount of segregation. And so the system will be much

likely to be stable under those conditions. So in shelling, it's literally a checkerboard. Even in my book, the big picture, I talked about the shelling model a little bit. And so it's literally your nearest neighbors, which makes sense if you're talking about racial segregation in a city or something like that. And so you're saying you're putting it on a network, basically, right?

Where there's different nodes that you can hear, Or is it more dynamical than that? Is it just like a different spatial structure or something that it changes with time? I focus on groups in this case, so more like subredits. So it's more like you join a community and then you're exposed to random people within that community. You can also run it on networks. But in this sense, the network structure are less prone to this emerging as kind of shilling dynamic because you would have to, you need a kind of transitivity.

You know, you need something like if you're leaving the community, the community becomes more segregated, and that increases the chance of someone else moving. So you get this kind of threshold effect where, you know, one person triggers another person. And in the network, you have to make really strong assumptions for that to be the case. Oh, okay. So if you get annoyed with someone, you just unfollow them. But that doesn't change your, you know, your friends networks. And so they're not going to be more likely. So there's not the positive feedback you get in your, in what you did?

At least not this kind of shelling feedback. Good. Okay. And by the way, my impression is that, you know, Schelling was offering an explanation for urban segregation that did not require, like, you know, racism handed down from on high via redlining or whatever. It was all just individual preferences. But in fact, when the social scientists have gone to look at it, the reason why real cities are segregated is, in fact, because of racism from on high, forcing it to happen. Yeah, no, I think this is a really important point. And it's quite funny in some ways.

I've been in geography for not anymore, but I was in postdoc in the geography department for about four years. And quickly realized that the only social scientists that do not really know or engage with the shalling model of segregation are the geographers. Because it just seems fundamentally incompatible with that way of thinking. And I'm very much in agreement. And it is quite interesting to a certain degree. And I think it connects to the question of epistemology. because it's kind of the Thomas Schelling segregation model, it gives a very deep insight into the kind of dynamics of segregation.

But it's also really hard to kind of bring that insight into dialogue with the existing literature on segregation in cities that, as you say, very much, point to kind of structural racism, redlining. But to a certain degree, I mean, both are true, right? Yeah, exactly. It's just difficult to make these theories kind of speak to each other. Well, my line has always been that the Schelling model is really good at explaining exactly what you started with, which is where people sit in the cafeteria, right? Like, there's not rules like, you know, the jocks and the nerds have to sit on different sides, but they always do because of exactly these preferences.

Yeah, no, I think that's a good point, and it's probably also less of a provocative example than using it to think about our segregation. And so this idea that, you know, we do change our social network or social media usage to be just a little bit more within a set of people that we want to hear. Like, where does this apply in the real world? Is this, are we thinking of Twitter or YouTube or TikTok or Facebook or what? Yeah. So this has basically been a long debate in the social science is this kind of question. of how pervasive echo chambers are or not. And it's still a very, very heated kind of debate.

But what I would say is basically that there are suggestions that there is quite a lot of communities that are relatively segregated. And so looking at, for instance, most subredits are, if they are political, they tend to be towards one side or the other. But I do think it's an interesting. I mean, Twitter has historically been a good example of the opposite. It was for a long time quite inclusive in the sense of having both political sides. And it's an interesting because Twitter has kind of functioned as the kind of model organism for social science research for looking at platforms because it's been one of the few platforms where we can actually get a lot of data or we could.

And so a lot of kind of computational social science research has looked. at Twitter and used it as a kind of way of speaking about social media. I would say that Twitter is like it was a very different platform from everything else because it actually had all political sides and it was characterized much more by a kind of conflictual debate. But if you look at kind of smaller communities, they do tend to be much more segregated in terms of opinion. And that can be a problem in the sense that if, you know, political theorists when they talk about what conditions need to be fulfilled for us to have a kind of functioning political discourse, functioning deliberation. One of those conditions are that we need to have kind of diversity of opinions.

We can't just have like political side, which is, I mean, pretty obvious, I guess. I guess I was going to ask a question about that. How bad is it if people on social media interact with people who are like them? Like I can imagine maybe a utopian political structure wouldn't be like that. But, you know, most people on social media are not there to be utopian political actors. They're there to talk to their friends and be reinforced. Is that so terrible? So, I would say that I think in a lot of cases, it can be even very beneficial.

I mean, in some way, one of the like ways that social media transforms, society, one of the key ways, was this possibility that we couldn't connect with anyone from all over the world. And so for a lot of communities, especially in minorities, you know, if you're LGBT and you grow up in a small village somewhere and you don't, you know, have anyone can connect to, it's been shown that it's very beneficial for your mental health and for your experience, for your lived experience. at the same time, the way that it affects politics is not always as beneficial. Right. Because obviously, if the minority that you belong to happen to be kind of, you know, some extremist form of, you know, neo-Nazism, it seems to have similar kind of consequences for those communities because it allows them to come together and form a kind of shared sense of community.

and it transforms them from being, you know, someone isolated to a kind of confident political community. And that can be quite dangerous in terms of radicalization. And is this, I mean, my impression is that it is something that comes from newfangled technology, social media, things like that, the ability of these smaller groups to come together. Like some of them are just going to be people who like to crochet and others are going to be neo-Nazis, right? But is there data that backs that up in the sense that have we seen more viability of these small groups than we did in the 1960s or whatever? I mean, it's very difficult to kind of look at those kind of changes, right?

Because we unfortunately, we only have one society. It's hard to compare how society would look different without social media, without digital media. But what we can say is that we've seen a kind of increase. in political violence. We've seen a kind of democratic backsliding in a lot of countries. And we've seen the kind of political extremist movements entering to the political mainstream. And whether or not that is causally linked to social media is very difficult to say. But it is very clear that it is in our current society very much entangled with social and digital media.

And I've looked, so in my previous book, intimate communities of hate, we look at one of these online communities. And basically we try to answer this question by going in depth and looking at the stormfront community, which is a very old kind of Nazi community in the U.S. They predate social media, right? Yeah, so basically it goes back to like 95. And the nice thing about it is that all of the data, all of the conversations over this, you know, long period of time is all available online, you know, if you're able to scrape it and bypass their very securities from preventing from scraping it. So we have all of the data and all of the conversations over this, you know, 20 plus year period. And so that allows us to kind of look at how the users are changed by interacting with this community. So we can kind of use natural language processing and various forms of text analysis and kind of see of how individuals, when they interact on this community, how does it change their language and different markers of like how they perceive themselves and so on.

And the kind of image that we come out with is much more, you know, it's a very much kind of a question of a community formation of changing identity and so on. And so just an example, we can kind of see how when they first come in, they use. I and my and speak of themselves. But then over time they start saying we or SF, you know, for Stormfront, because they start, you know, that's kind of a marker of them starting to think of themselves as part of a collective and as part of something larger. We shop for everything at home now. Why can't we shop for blinds at home too?

That's why I love three-day blinds. Instead of wandering around a store, three-day blinds sends a professional design consultant straight to you. They bring all the samples, help you see what actually works in your space, take expert measurements, and even handle the installation. We've all bought something that looked great in a store and looked terrible once we got home. With three-day blinds, you see everything in your actual light, with your actual furniture, before you commit. Plus, whether you're home or away, you can control your window treatments with the three-day blinds app.

So for a free, no charge, no obligation consultation, head to 3dayblinds.com slash pod 50 for our buy one get 1 50% off deal. That's the number three, DAYBlinds.com slash pod 50. So that's very interesting because it's not just about people in the study we were just talking about, you're talking about associating with people who are like-minded. And here you're talking about the feedback acting on yourself. and the individual people are sort of changing their identities in response to those interactions. For sure. I mean, I think it's very clearly, you know, kind of a feedback process, right, where we have a kind of segregation mixed with a kind of changing of identities. And so we can kind of see, you know, in the Stormfront case, especially after the 2008 Obama election,

the few days after the election, there was just a huge search of users coming in. So new people joining and looking at what they're saying, we can see this kind of emotional kind of confusion and anxiety. And they're trying to, you know, they somehow feel confused about this world that they're living in where, you know, a black person is like articulate can become precedent. And what does that mean for their self-identity and like how can they make sense of this? And then the community function is kind of, you know, emotional talk therapy that allows them to find new narratives. and kind of resolve these emotional anxiety and turn it into kind of from something passive like anxiety to something active like anger or outrage. And they come out with these narratives about, you know, there's absurd narratives about the Jews and, you know, like the actually the whites are the superior race. It was just, you know, whatever that happened.

And, you know, so it's very much a kind of process that is on the level of like identity and emotion. and it's kind of self-narratives. It's very interesting because it all cycles back with the very first podcast episode of Mindscape. I interviewed Carol Taveris, who's a social psychologist, and she has this idea of the pyramid of choice when, if you imagine two people who are basically 50-50 as to how they could make some certain choice, right? You know, what sneakers to wear or whatever. But once they make the choice, if they make it in different directions,

they start justifying that choice to themselves. And they end up very far apart, even though they were essentially. indistinguishable before they, you know, collapse their wave function on that particular option. Yeah, no, that makes sense. And so, I mean, ultimately is kind of the question of the kind of structural context in which people are interacting that can produce this kind of outcomes. And so, yeah. Can I ask about either your model or the Schelling model?

There seems to be, like, as the physicist in me thinks of the Izing model, which is similar but not exactly the same, where you have spins that are interacting on a lattice. And the thing we do there is we introduce probabilities by having a temperature, right? There's some chance that the spin is going to flip and whatever. And in the Shelley model, there's a probability because when the person decides to move, where they move to is random. But the choice about whether to move or not is not random, right? That's just determined by how many neighbors they have of each kind.

So have people done that? Have people introduced a probability of moving rather than a certainty and seen if changes anything? Good question. I honestly don't know. Okay, but I don't know either. Someone should do that. Someone listening out there. Maybe something for us to do. Yeah, yeah, absolutely. And then, okay, so you then did a different study, which came out also very recently, using large language models. And here, well, I'll let you tell the story, but the idea is rather than just having these mindless dots on on a grid or whatever that are interacting with each other. You literally had little agents talking to each other and making choices in a social media context.

So how did that go? Yeah. So basically the aim of this is to address, in part, this like kind of longstanding criticism from social scientists or at least from a lot of social scientists when it comes to agent-based modeling, which is that, you know, these rule-based agents are just not very good representations of the full spectrum of human behavior. which is, I think, fair enough. And I think, I mean, in a lot of cases like the shell and segregation model,

this simplicity can be, it is very useful and it allows us to, you know, throw light on some emergent phenomenon that is ultimately structural. But in other contexts, it is also limiting, right? So looking at social media, like politics on social media, for instance, it's an example where these richer behaviors, they also, can really matter, right? We cannot really separate the cultural from the structural. We need to look at them as kind of intertwined and as interacting.

And so that's kind of the background, because what I'm interested in here, what we're interested in is this question, because we basically, you know, we spent 20 years or something criticizing social media and pointing to the problems and a link to various problematic outcomes. But now there's more and more kind of interest in it, like, can we be a little bit more constructive can we actually like do something about this because ultimately if social media can shape a politics that is like outrage driven and radicalizing it should also be able to shape a form of politics that is you know pro-social that has like healthier political and social outcomes yeah

so that's the kind of idea but and how do we study that well using observational data doesn't really work so that's kind of modeling approach can be really beneficial and so we're using agent-based models but having instead of these rule followers, we're using large language model. And they work as kind of stand-ins for humans. Maybe if I could interrupt you just quickly. I mean, maybe give a little bit of background onto the concept of an agent-based model, like as opposed to what? What kind of models are not agent-based and what are agent-based? And what is that used for? Sure. I mean, so in the social Sciences, the way that we have traditionally approached the social world is to link back to where

we started very much as a kind of complicated system. So we tend to think of society as kind of variables interacting, which in a lot of cases can work really well. But if you're thinking of this kind of complex aspects of the social world where you have interaction between agents and then leading to unexpected outcomes, those traditional kind of variable-based approaches just don't work at all. Like how would you like using used variables, how would you like study the murmuration, you know, like of birds? It wouldn't be possible. And so agent-based models is kind of using this kind of bottom-up modeling approach where we and so one example would just be the shelling model, right? So you have agents that are individuals, they follow simple rules and then you look at the outcomes.

And that allows you to think together the kind of micro behavior of individuals with system level outcomes that can often be kind of unexpected from the rules that you're coming in with. So the individual agents need not be very complex themselves. Traditionally, they haven't been. So they traditionally have been kind of simple rule followers. So you have like the shelling threshold rule or you have maybe an optimization rule. But basically building agents that would kind of mimic human behavior. in terms of reasoning or language production, it just becomes impossible traditionally, right?

Like, it would just be extremely complicated. You would have to build a kind of, you know, a reasoning agent, which we just haven't had up until, you know, a few years ago. But so basically, when Shachip-T came out and with the kind of rise of large language models, it just became a huge amount of interest in like whether we can use these models to, as part of agent-based models,

to kind of simulate social behavior. And so that's the kind of what we're doing, but we're trying to actually use it to contribute some kind of social scientific theory and contribute to our understanding of social media and its dynamics. And so what exactly was the experiment you did? I think of it as roughly speaking, letting lose a bunch of LLMs on a fake social network. That's pretty much it.

But basically our idea was to try to create a social media, platform and make it produce the negative outcomes that have been observed on real social media, and then try out a bunch of suggestions from the literature on how we can address those problems. And so our expectation coming in was kind of that we would have to, you know, fiddle a lot with the system and like try to make it produce problematic outcomes. And then so we could then see like how stable those outcomes are and how easy, like what's kind of solutions are best for a the problems. And so the problems that we focus on,

it's kind of conditions of social media that make public deliberation or public conversation difficult. It makes it difficult to have a kind of a functioning politics playing out on these platforms, drawing on kind of political theory. And so there are three different things that we've already touched on a little bit. But so one of them are echo chambers that you do need to have, If you're going to have a kind of constructive conversation across the political divide, you need to have both sides of the political divide present.

Otherwise, it's going to be really hard. So that's one condition you need. And then the second is this kind of question of attention inequality that we're also touched on. If you're going to have functioning political discourse, you need to have relative equality among individuals. So you can't just have like two or three individuals dominating the entire conversation because that's not a public discourse. That's just broadcasting.

And then finally, what's been referred to as the kind of social media prism. So this is the idea that you need to have a kind of constructive debate where people are actually trying to come to a solution. And so that speaks to this question of that social media has kind of tended to benefit loud, polarizing, conflictual voices. And that is very much kind of undermining functioning conversations. And so those are the three outcomes. that we were trying to kind of see if we could produce. And we were expecting that to be quite hard, to be honest. That's so sweet that you thought that would be hard.

Well, I mean, the literature has kind of pointed to or argue that a lot of these are, especially the kind of the social media prisms or this kind of the polarizing tendency, that those will be expression of engagement algorithms and that they would be like the expression of social media, identifying the most outrageous things that are being said, and then shove it in your face to kind of make you upset and increase the probability that you will comment or engage with the post. So, sorry, so the sort of two alternatives that we're trying to test here are, one is that when you get these echo chambers and polarization, things like that, it's the algorithm's fault

or the platform's fault versus this is just human nature. Well, so I'm not sure if I would put that as the context, really, because it's also like in this study, we don't, you know, I was honestly just kind of assuming that it was from the algorithms at least, that it's not just something that would emerge from human behavior. Right. But these are kind of structural outcomes from the interaction between people and the rules of the platform.

But to me, at least this kind of social media, it's such a, it's a rather specific thing. It's kind of odd outcome that the most extreme voices get more attention. And so I was expecting that to be, and I've written about this before arguing that is the kind of what I've called it trigger bubble. So, you know, it's not the filter bubble, but the trigger bubble. It's the social media algorithm trying to trigger you, make you upset in order to make you engage, because that's how the platforms ultimately make money. they make you post something, they draw information,

they figure out who you are and they sell ads. And so that was kind of my expectation. But basically what we started was just building the most, you know, bare bones platform we could imagine, which is just the agents. So I should say also that the agents, their personalities are, we take the ANES, so the American National Election Survey, which has very detailed information about

but US citizens, including their politics, and like, you know, if they like to go fishing, you know, we have a very detailed information. And basically, we turn that into a description for persona because, you know, LLMs, they really love to impersonate people. You know, you can ask it to, you know, explain your microwave and the voice of Shakespeare and it will do an excellent job. So just to be clear, so the LLMs that you let loose on the social media, they weren't all, you know, Adam and Eve. They weren't all tabular raw. from the start, you gave them a backstory. Exactly.

And the point here isn't necessarily that, you know, we want them to be completely realistic encapsulations of, you know, of this particular individual that they're enacting. We just want to have kind of a diversity and want to capture a little bit of the, a little bit of that cultural richness, you know. And so then they are allowed to interact on this platform. And the platform is very simple. So basically they can look at the, they see the most recent news,

so random selection of news from the specific day that we're simulating. And then they can choose to post about these news and they can just write a post about whatever based on it. Or they can see the timeline and they can choose to repost what someone else has written, someone that they follow. And they can also see, based on the post that they see, they can choose to follow someone that shows up in their feed. And if they follow someone, they can follow someone,

they kind of go into their timeline, they see their little presentation about them, and then they see their most recent posts. Wow. Okay. And can they unfollow people too? No, we don't have unfollowing. It's really bare bones. Okay, but there is some ability they can choose to follow someone or not. Exactly. And so what we're looking out at the outcomes here is the network structure that emerges through this interaction. Yeah. Because we want to have simple measure things and we don't want to just look at, you know, how they're talking or something that's like an immediate outcome of the large language model.

We want to have something that's more an expression of the structure of the platform. And network structures are interesting in the sense that they are very much kind of emergent and they are produced through their structural outcomes. And so that's what we're focusing on. And so we're trying to identify these three attributes. And to our surprise, we didn't actually need to do anything more than provide. this bare-bone platform and we got these three features that are that widely consider the problematic aspects of social media. But I should say that that doesn't mean that engagement

algorithms aren't problematic. That doesn't mean that, you know, that, you know, it might still be that they're making matters worse, but it does imply that removing them will not completely solve the problem. So how many users did your social network have? We ran it with 500 users. 500, okay. Which is, it's a little bit of a limitation with the approach. It's that it is quite expensive compared to running a conventional agent-based model. Right. And conventional agent-based model have always been criticized

for being very expensive to run competitionally. So it is kind of a weakness of this approach. We shop for everything at home now. Why can't we shop for blinds at home too? That's why I love three-day blinds. Instead of wandering around a store, three-day blinds sends a professional design consultant straight to you. They bring all the samples, help you see what actually works in your space, take expert measurements, and even handle the

installation. We've all bought something that looked great in a store and looked terrible once we got home. With three-day blinds, you see everything in your actual light, with your actual furniture, before you commit. Plus, whether you're home or away, you can control your window treatments with the three-day blinds app. So for a free, no-charge, no-obligation consultation, head to three-dayblinds.com slash pod-50 for our buy-one get-one 50% off deal. That's the number three,

d-a-y-y-y-blinds.com slash pod 50. Pasadena Playhouse presents learner and Lois Brigadoon, featuring six-time Emmy Award winner Tyne Daily and a full 21-piece orchestra. With its a lush score, including Broadway standards like almost like being in love, and the Heather on the Hill. This beloved classic is a heart-stirring journey

into a world where times stands still and love device all logic. Step through the midst of the Scottish Highlands into this timeless love story. Learner and Lowe's Brigadune is on stage at Pasadena Playhouse beginning May 13th. Tickets are on sale now at Pasadenaplayhouse.org. And you said that the bad outcomes happened.

Reminded us what the bad outcomes were. There's a list. Yeah, so basically you get echo chambers. So Democrats and Republicans end up not following each other. They're just talking to themselves. You get high levels of inequality, basically power law distributions of attention. So a few users kind of dominate the entire discourse. And then finally, you get what Chris Bale has called the social media prism,

which is that the more polarized, more extreme users tend to have more attention. Okay. Okay. Okay. And is there, I mean, I, it's very, very tempting to speak anthropomorphically about LLMs, right, and to ask about their motivations or something like that. But of course, that's, you know, illegitative. And that's just faking that. They don't really have motivations. But is what is the right way of phrasing the answer to the question? Why doesn't LLM want to follow people on its own political side of the spectrum? So, I mean, so where these outcomes are kind of stemming from, like, I think it's a little bit different outcomes each one. But so for instance, that we get the power law distribution, that we get very unequal distribution of attention. I mean, that's kind of stems from preferential attachment, right? The probability of you getting a follower is proportional to how many followers you already have. So that, I wasn't so surprised to a certain degree that that emerged because it is a well-known feature of network, as I,

as I already mentioned earlier. But the other kind of dynamic that we identify that we haven't really seen studied before, because you do kind of need to have this kind of special large language model network combination to be able to study it, which is the fact that basically we know that retweets, people repost or sharing is very effective,

it's very emotional, it's very much reactive. So we see something that we're upset about or that we're really emotional we're reacting to. And those are the types of things that we tend to share. And that is well known. And it's been kind of argued that that shapes the content that we see on social media. But what we're adding to that is that it's not only shaping what kind of content you see, but it's also shaping the kind of construction, the gradual formation of the network structure. And so you have a kind of feedback effect between what is shared, which is very much emotional, and who you follow.

And that is what's creating this kind of dynamic where more polarized users tend to have more followers and get more attention. Because it is the kind of sharing is part of this kind of feedback process. I understand how an LLM can respond to a query and say something. But did you have to cook up an extra set of instructions for when to follow somebody? We only, we don't provide them with any, you know, it's just like this is your persona. how would you, based on this persona, how would you act in this situation? And then just have them kind of respond on the basis of that. Okay.

And the persona includes their political orientation as well as whether they like fishing or whatever. Exactly. So it contains their political affiliation. And so in terms of the echo chambers, it's also not maybe so surprising. And we can also look at the kind of motivations that the LLM is giving for wide issues to follow them. And it does, like in terms of the echo chambers, it's kind of like, okay, so. I don't want to, like, since I'm Democrat and I feel strongly about it, I don't want to engage with this person who's from the other side.

But I think it is interesting that it goes beyond just that individual choice. It's not like the dynamic and the emergence of the structural echo chambers is not just because individuals are choosing it, but because their choices are impacting others because it shapes what messages are shared and how the structure of the network is gradually built up. And is there room for the individual agents, as it were, to be affected by their success on the network? Is their audience capture? Do the LLMs learn to be more provocative so they get more retweets? Not in, so this is something that I think is really important and I think it's really central to how social media is reshaping politics because it's not just that we get more polarized when we go on the platforms, but the platforms are kind of shaping the incentive structure. that are, you know, shaping society because they are defining who gets attention and who doesn't.

And so in that sense, it's reshaping a politics built around being kind of outrageous and so on. But no. Okay. Briefly put, no. But this is like a paper that I'm currently working on, actually. Okay, good. Very much what I'm interested in. And of course, someone is going to say, look, LLMs are not people, right?

Like what worries about limitations of mimicking human behavior should we have in the forefronts of our minds when we're using these LLMs? Yeah, so this is, it's a really kind of big debate right now because there's a lot of kind of excitement around this kind of generative simulation. I've tended to be on the more skeptical side of this. So me and my co-author, Mike Leroy, we also have another paper coming out where we were specifically trying to, to answer this question and like how much can we trust this and how much can we trust that they're valid as representations of human behavior. And it is kind of, it's a very interesting question to certain degree because the models are more realistic as representation of human behavior than just a list of rules. But they're also much harder to validate, right? Because you don't know exactly

what is playing out and it's very hard to calibrate their behavior to match human behavior. And so it becomes kind of like a difficult question. They're like ending up somewhere in between empirical methods that have a high level of validity and formal models that are parsimonious that are easy to understand and they're neither and so it's a little bit unclear how we can use them and the way that we think about that in this paper is precisely, it's linked to the fact that we're not looking at the text that they're producing. We're trying to look at kind of structural features that are emerging as aspects of the platform not that you know we're trying to distance ourselves from their behavior

and try to look at the kind of structural outcomes of that behavior. And to a certain degree, what we're interested in is the kind of the robustness of those outcomes. So in the sense, what do I mean by robustness? Well, basically, similar to the Schelling segregation model, you can basically, it's the kind of joke among models. Whenever you build a geographical model that has any type of lattice in it, you always get the shelling segregation effect. It's just such a stable outcome. and you have to really work hard to study any other effect. And so that's the kind of effect that we're interested in.

We're interested in seeing how robust these emergent outcomes from the platform are. And so if we were just looking at the kind of toxicity of their conversation, I would feel less confident that these are something that are actually playing out in real world. But given the robustness of these emergent patterns, I have more confidence. But that being said, we definitely also need to do more studies on this. Well, it does seem like the robustness from shelling to your LLMs. I mean, these are sort of different individual small-scale dynamics giving rise to similar, large-scale behavior. It sounds like there should be a theorem, like the second law of thermodynamics,

but about polarization or segregation or something like that being the state with lowest free energy. I don't know what to call it. But are there theorems like that in the statistical mechanics of social media? So there are people doing these much more, you know, physicist approaches, kind of opinion dynamics. I have a colleague here who is basically doing icing models wrapped into icing models in order to understand social media. my approach tends to be a little bit more based on empirical data and trying to be closer to the system. So I think it generally is hard to find, unfortunately, this kind of loss in the social world because it is such an open system. And I mean, linking back to the question of whether society is a complex system or a complicated system,

I would argue that it's neither or it's both to certain degree. And one of the features of polarization that is a little bit weird to me in the United States is how even it is. Like we've essentially always in the history of the USA had two political parties with roughly 50% representation each. It never goes to like 70, 30 or anything like that. That's not something you can test your model, right? I mean, you basically baked in into your initial conditions, how many, Democrats there were and how many Republicans? Yeah, no, so this is not something I would capture in this model.

In general, I'm thinking of these models as just trying to somehow capture, you know, one mechanism. Yeah, sure. And as soon as you start having a lot of mechanisms, it's, it's very, very difficult. So in certain sense of like how the outcomes of this interaction feed back into society and how it drives polarization, that's to a certain degree outside the kind of bounds of this model. So you said that you were tracking the statistical results, not the individual words that the LLMs were putting into their posts or whatever. But do you know that those backstories you gave them? Like, you know, this person lives in Boston, you know, they enjoy theater or whatever.

Did that matter? Did that affect how they sorted or polarized or succeeded in the social media game? I mean, to a certain degree, yes, because if I wouldn't have given them any sense, sense of if they were Democrats or Republicans, they wouldn't be able to act as that and you wouldn't, you know, trigger this kind of feedback effect. But if it matters that they, that they know if they go fishing or not, that probably doesn't matter. Well, I don't know. I would be curious to know. I think of that more as a kind of a little bit of noise or a little bit of peribation to make them not only act on the basis of their political personality, but that there's also like,

okay, so this, you know, he likes fishing. You know, he's talking about fishing. I like fishing. I'm going to follow him. So it adds a little bit of that kind of noise and the fact that, you know, our lives are not just politics. It's, that's just a small part of everything that we are. We have much more rich identities than that.

And you injected news, basically? Is that like, you know, there was an external source of perturbations that said like, you know, this news event happened or whatever? Yes, exactly. So basically, if you're just giving the agents, if you're just letting them talk without, you know, having them something to talk, giving them something to talk about, it just becomes the most generic. Not interesting conversation. And it also doesn't create this kind of richness that, you know, also functions as a kind of noise, you know. And so we basically focus on a certain day and we got all the news from that day.

And then we present them, we like a random selection of those news and then have them discuss it. Okay. So you didn't totally make up the news. were inspired by real news. Exactly, yeah, we got real news from a particular day. And so I guess you mentioned the power law distribution of attention. So some of these, they're all LLMs, but some of them get a lot more followers than others. Is there any sense in which some of them are just better at social media than others, or is it purely statistics and randomness? So I would say that it's pretty much stochastic. I mean, I,

I wouldn't say that it's just randomly, one of them happens to be a great influencer as such. But being more political and being more extreme does help to become more influential, but it is, I would say, pretty much stochastic. And I mean, that fits also. There's been various like kind of experimental studies on on these power law distributions and how they can emerge on in systems just through the feedback effects. and it doesn't need to be any difference between the things that are being selected, you still get these power laws.

And it's just kind of random. So did we learn anything about how to make the world a better place through doing this? Does this help us suggest any ways to make social media better? I mean, so I guess we didn't really mention the interventions that we tried out because we built this platform and then, you know, we saw these negative outcomes. And that gave us the baseline where we could try, okay, can we fix this problem? That became the kind of next step. And so we looked at the literature and looked at what has been suggested,

what are people kind of optimistic about in terms of trying to solve these problems. And basically we had the kind of wide variety of more or less sophisticated solutions that have been presented. One of them was kind of the bridging attributes, which Jigsaw has released, which is a subcompany of Google, which is a subcompany of Google, which is basically they analyze the content of messages and then you can sort your news feed if you're a social media platform you can use it to sort your newsfeed and you get the most kind of constructive comments

those are the ones that you show so instead of you know like instead of showing the most upsetting comments you can show the most constructive the most partisan so it's called the bridging attributes so that's one example And another example was like, is this smaller solutions, just like hiding the biography, the little description of the agents when they follow each other. So they don't know if the other person is Democrat or Republican, for instance. And another is just sorting chronologically instead of, you know, instead of showing the most

shared posts. So basically we try out a bunch of those solutions that have been suggested. But I should also say that we are doing like fairly. extreme versions of it that wouldn't necessarily be realistic to implement on the platform. So for instance, we show the, you know, one algorithm where we show the least liked posts first, which is probably it would lead to a really awful platform if you implemented in the real world because we want to see, you know, like the most extreme solutions kind of. But unfortunately, none of these solutions really fix.

the problems that we're observing and some of them actually make matters worse. So, for instance, the chronological timeline actually leads to more of a social media prism where you get more extreme users get even more attention. And so it seems, what we take from this is that these, this kind of emerging phenomena seems to be very rigorous to perturbations. That is basically a little bit like this. shelling segregation effect is a very robust emergent phenomenon. Yeah, and you don't want to have a social media network that just tells every user who to follow, right?

You need to give them some agency there. Yeah, I mean, it's also the question of if people are going to use the platform or not. But basically, I mean, to me, what this suggests is that this basic structure that we see across social media platforms where you have a network, you follow people and you repost things, that that tends to be linked to these problematic outcomes. Well, I guess that was where I was going to go. You already sort of said this is hard to study, but is it something truly new, these social media things?

We used to be happy just getting the conventional news on one of the three network stations on TV, and now we have a lot more variety in what we can listen in on. but has this had a big effect? You can see that it has an effect, but I guess how much of our current political mess can we be tracing to this? I know it's a hard question to answer.

Yeah, I mean, it's ultimately impossible to know. And to a certain degree, it's also a bit tricky treating social media as something that's like external to society and that happened to society. Because to me, that social media is structured the way it is, is very much an expression of, you know, coming back to Fordism, the kind of transition from an industrial society to a post-Fordist society where the focus of capital is to, is advertising and figuring out information about you,

that it was not catering to a mass market where, you know, everyone, you're selling the same product to everyone, but that you're really trying to not only identify consumer niches, but even create consumer niches. And that, like, that, basic fact of how the companies make money, the business model underlying social media and the internet, that has very much shaped what social media has become. Of course, it's also feeding back, you know, but it's very difficult to say how social media would be different if it wasn't

that context. Well, I guess it's the feeding back. I was going to mention very briefly, like, it's not just that you have social media in addition to mainstream media, but the social media affects the mainstream media. They want those clicks too. Yeah, exactly. I mean, I think this is something that people often mention as like, okay, but we can just, you know, I can just stop using social media and I won't be affected by these negative consequences. But of course, that's not the case, right? Because social media is, as I mentioned, it's reshaping our politics and it's very much reshaping also mainstream media. So a student of mine in a student project in my course. Last year, for instance, he looked at the New York Times headlines over time and then measured how click-baity they are.

And basically what he saw was that when social media entered on the scene in 2010, and so you saw a kind of jump and you saw that the New York Times also changed how they wrote their headlines. And I mean, that's just the kind of, there's one expression of it, but of course it's reshaping, you know, the incentives of attention produced by these platforms are, reshaping our politics, our media, and our culture overall. And what does clickbait mean? Is it a function of sort of giving less information and saying, like, you won't believe what happened next? Yeah. There's like actually a bunch of kind of features of text that make them more or less clickbaity. But basically, the way he did it was to just

look at the databases of clickbait news articles and then non-clickbait news articles and then train a classifier on it. So I guess I'll Last thing to talk about is there's polarization. So you know, you had these LLMs. They sorted themselves in a shelling-like way, et cetera. Can we say anything about the quality of the information, like the truthfulness versus misinformation? Our social media helping us not just only talk to people like ourselves, but to get it wrong by sharing miss and disinformation?

So I would say, I mean, this is not something I'm looking at in this specific model, because the in part because you're the LLMs are under you know open AI doesn't let them produce misinformation so you can't really use them to study that but I have looked at at this in the context of using actual social media data and I mean what I would say broadly around it is that social media is by removing the kind of gatekeepers that we used to have from mainstream conventional media they're also and and by creating kind of really strong incentives for gaining attention and shaping the kind of conditions.

But they're really producing conditions where not anchoring what you're saying to truth becomes a kind of beneficial strategy for gaining attention, right? Because you're both, it allows you to be kind of outrageous, it allows you to trigger people, and you're not really constrained by reality in the same way. And of course that also becomes interconnected with politics. So I had a paper coming out with my co-author, Juliana Schweri, earlier this year,

where we look at politicians across countries, and we look at their Twitter posts over a five, six-year period, and we look at all the examples of when they're shared links, and we identify misinformation through that. And so we can link each politician to their likelihood of sharing misinformation. And then we can basically use that for a kind of comparative model. So basically a statistical model to identify the conditions when politicians are spreading misinformation. And so this links to this broader question of the link between social media and the spread of misinformation,

which has been a big debate around this, especially in the last few years, about whether it's just social media reducing the quality of information overall. Right. And what we argue is basically it's not just that, but that social media becomes intertwined with politics. It becomes intertwined with different political movements. And the result is that certain political movements are emerging, shaped by the interests and incentives of social media in such a way that they use misinformation as a political strategy to gain advantages in political competition. And what we find in that study is basically that it's specifically the, radical right populist parties that are driving this rise of misinformation.

So it's not just a social media phenomenon in itself, but it's social media intertwined with politics and political systems. So I'll let you give like a last big picture kind of thought here. Like am I getting the impression that social media are just bad that it was a mistake, that their net effect is negative? Or can we have some shred of optimism to hold on to? I mean, I think that there are. are also kind of some degree positive outcomes from it.

And like for certain communities can be beneficial. And I mean, to a certain degree, like growing up, I loved the internet. It was great. I grew up on this, you know, in the countryside on this old island in Sweden in the middle of nowhere. And like the internet kind of provided a social world for me and allowed me to, you know, connect to ideas and everything. So and to a certain degree what I would hope for is also kind of kind of, of going back to that, you know, innocent era of the of the 90s, of the internet of the 90s.

Like, you know, I was using ICQ and you had this button where you could click and talk to like a random person anywhere in the world. And I just like love that. I spent my days like talking to some, you know, some random person in Arizona. And to a certain degree, you know, I mean, of course, that was like a more innocent time. And maybe if we would try to bring that back, it would, you know, lead to something horrible these days. But I do think that, you know, we could create structures, like we could create platforms and spaces that would actually be, you know, beneficial for us and that would actually be positive. It's just we might need to rethink it in more fundamental ways than just this cosmetic changes to algorithms or designs.

All right. That is something, there's homework out there for all the young people. Think about, you know, fundamental changes we can make because they're not going away, you know, even if social media or a net bad. They also are good and we're going to have to live with both of them. So Better Torenberg, thanks very much for being on the Mindscape podcast. Thank you so much. This was really great.