(ALMEDIPS) Algorithmically-Mediated Digital Public Spaces

Algorithmically-Mediated Digital Public Spaces

The study of the consequences and the explainability of AI recommender systems impacting complex social systems such as digital public space (DPS) has become a relevant stream of research. Part of this interest has been spurred by hypothesized disorders of digital public spaces, such as polarization and extremism, that might be linked with algorithmic recommendations. The policy community is quickly developing around the measurement and regulation of AI systems with high societal impact, including those acting on DPSs1. One difficulty in evaluating AI systems in DPSs is the abundance of case studies, metrics, and sometimes contradicting results. This is explained in part by the lack of spatial models on which to operationalize notions that appeal to geometrical settings (distances, directions), such as extremism (distant individual opinions) and polarization (groups with distant opinions). Web traces have been traditionally analyzed with graph models lacking geometrical notions. Today, there are no widely accepted definitions for these phenomena, nor ways of measuring them. This makes it difficult to determine the role of AI systems in fostering these states, let alone designing AI systems capable of mitigating identified disorders. This project proposes to tackle this by Obj1) developing dual topological-geometrical representations of DPSs modeled as a heterogeneous network of users and entities (articles, videos, groups, etc.) embedded in multidimensional opinion spaces on which dimensions stand for indicators of attitudes towards issues (e.g., European integration, immigration, authoritarianism) using classical political sciences datasets. In this model, where polarization and extremism take operationalizable meaning, this project proposes Obj2) to collect and to simulate recommendations (using methods form the state of the art) to assess the role AI systems in these disorders.