Ars Technica: I’m skeptical of AI in general, particularly in a research context, but there are very specific instances where it can be extremely useful. This strikes me as one of them, largely because your basic model proved to be so robust. You got the same dynamics without introducing anything extra.
Petter Törnberg: Yes. It’s been a big conversation in social science over the last two years or so. There’s a ton of interest in using LLMs for social simulation, but no one has really figured out for what or how it’s going to be helpful, or how we’re going to get past these problems of validity and so on. The kind of approach that we take in this paper is building on a tradition of complex systems thinking. We imagine very simple models of the human world and try to capture very fundamental mechanisms. It’s not really aiming to be realistic or a precise, complete model of human behavior.
I’ve been one of the more critical people of this method, to be honest. At the same time, it’s hard to imagine any other way of studying these kinds of dynamics where we have cultural and structural aspects feeding back into each other. But I still have to take the findings with a grain of salt and realize that these are models, and they’re capturing a kind of hypothetical world—a spherical cow in a vacuum. We can’t predict what someone is going to have for lunch on Tuesday, but we can capture broader mechanisms, and we can see how robust those mechanisms are. We can see whether they’re stable, unstable, which conditions they emerge in, and the general boundaries. And in this case, we found a mechanism that seems to be very robust, unfortunately.
Ars Technica: The dream was that social media would help revitalize the public sphere and support the kind of constructive political dialogue that your paper deems “vital to democratic life.” That largely hasn’t happened. What are the primary negative unexpected consequences that have emerged from social media platforms?
Petter Törnberg: First, you have echo chambers or filter bubbles. The risk of broad agreement is that if you want to have a functioning political conversation, functioning deliberation, you do need to do that across the partisan divide. If you’re only having a conversation with people who already agree with each other, that’s not enough. There’s debate on how widespread echo chambers are online, but it is quite established that there are a lot of spaces online that aren’t very constructive because there’s only people from one political side. So that’s one ingredient that you need. You need to have a diversity of opinion, a diversity of perspective.