Periodic Reporting for period 1 - TWON (TWin of Online Social Networks)
Periodo di rendicontazione: 2023-04-01 al 2024-03-31
Experts, scholars, and leading political decision-makers warn that Online Social Networks (OSNs) have transformed public debate in harmful ways. Personalization algorithms, it has been argued, create so-called filter bubbles and echo chambers where users’ opinions are reinforced, amplifying processes of opinion polarisation. Despite frequent calls for interventions to minimize such undesired effects, there is no agreed-upon method for estimating the effects of changing the parameters of the design of a social network service. Crucially, the complexity of such systems makes it hard to translate results of isolated experiments into an estimate of the overall effects. The TWON project develops a novel empirical method for systematically researching the effects of design choices of mechanisms inside OSNs, by creating digital twins of social network sites, called TWONs. The TWON can then be used to study counterfactuals, such as: How would the effects look like, had the OSN been designed differently? In order to achieve that, the TWON project will combine empirical observations of existing OSNs, theory-informed simulations, and specific case studies. These form an iterative process, in which we will build and refine the TWON.
If successful, this would be a major leap towards a better understanding of platform mechanics, both for the scientific community and for societal stakeholders. The TWON project will produce evidence-based recommendations for regulatory innovations regarding OSNs and enhance digital citizenship by participatory methods. This can reduce the detrimental effects on democratic debates when platforms are primarily optimized for economic gain. TWON enables OSN research in a controlled but naturalistic environment that would not be possible relying on for-profit OSN operators. The effectiveness of the TWON method will be demonstrated in two case studies on two diametrically controversial debates: the conflict in Ukraine and health-related communication, such as in the case of COVID-19.
If successful, this would be a major leap towards a better understanding of platform mechanics, both for the scientific community and for societal stakeholders. The TWON project will produce evidence-based recommendations for regulatory innovations regarding OSNs and enhance digital citizenship by participatory methods. This can reduce the detrimental effects on democratic debates when platforms are primarily optimized for economic gain. TWON enables OSN research in a controlled but naturalistic environment that would not be possible relying on for-profit OSN operators. The effectiveness of the TWON method will be demonstrated in two case studies on two diametrically controversial debates: the conflict in Ukraine and health-related communication, such as in the case of COVID-19.
All of our activities centered around developing a very first version of a digital twin of online social networks. We did this in a modular way, such that components can be developed in parallel, but also changed and/or replaced by future, improved versions without compromising the overall infrastructure. In broad strokes, next to supporting work like collecting necessary datasets, performing evaluations, and so on, we did the following:
(1) We started development a "platform model". This involves creating a network generator that can generate different types of networks with different characteristics, which then can be used to simulate, for instance, different news feeds and different ways of users connecting to each other. In particular, we showed the importance of extending existing network models with "success-driven activity", illustrating that outcomes such as polarization change if previous success of agents is taken into account in a simulation.
(2) We started developing a "user model". In particular, we developed agents that, based on large language models, can react to content on social networks, according to different persona's (e.g. based on their political leaning or other features). These agents of bots can be linked up both to simulation models as well as to platforms with human users.
(3) We developed a social network site with a user interface similar to existing services in order to perform experiments with human participants.
(4) We designed a first study in which we had both human participants as well as the previously mentioned bots interact on this social network site. The first study was fielded shortly after the end of the first reporting period.
(5) We worked on developing metrics to evaluate debate quality, both from a theoretical/normative as an empirical point of view.
(6) We developed ethical guidelines for the use, design, and deployment of digital twins of online social networks.
(1) We started development a "platform model". This involves creating a network generator that can generate different types of networks with different characteristics, which then can be used to simulate, for instance, different news feeds and different ways of users connecting to each other. In particular, we showed the importance of extending existing network models with "success-driven activity", illustrating that outcomes such as polarization change if previous success of agents is taken into account in a simulation.
(2) We started developing a "user model". In particular, we developed agents that, based on large language models, can react to content on social networks, according to different persona's (e.g. based on their political leaning or other features). These agents of bots can be linked up both to simulation models as well as to platforms with human users.
(3) We developed a social network site with a user interface similar to existing services in order to perform experiments with human participants.
(4) We designed a first study in which we had both human participants as well as the previously mentioned bots interact on this social network site. The first study was fielded shortly after the end of the first reporting period.
(5) We worked on developing metrics to evaluate debate quality, both from a theoretical/normative as an empirical point of view.
(6) We developed ethical guidelines for the use, design, and deployment of digital twins of online social networks.
Our results are promising and go beyond the state of the art. While we are still in the first iteration, and hence need further improvements, we already see that our approach of integrating simulation approaches with LLM-based agents is a promising avenue to pursue further. Also, preliminary results show adding a user interface to enable participants to directly interact with our simulations is a viable and innovative way to balance the aims of collecting user data in settings that are as realistic as possible with having experimental control over the environment. We expect impact on future research designs, which we hope will take up our approach.