How do recommender systems learn political opinions? A semi-synthetic step-by-step experiment

Recommendations play a crucial role in shaping informational diets on social media, raising concerns regarding potential consequences such as political segregation. We take an algorithm explanability approach, as opposed to a description of recommendations, to show how recommenders inadvertently create geometrical representations of the ideological position of users with minimal and ubiquitous platform data, and how this impacts content diets. In comparison to work showing the existence of ideological structures in machine representations, we provide a step-by-step causal explanation of their formation. To achieve this, we compute synthetic recommendations with a model trained on real-world data from a panel of nearly 40 thousand X users and the contents they shared on the platform. We show that elementary recommendation principles trained on content dissemination data produce a spatial representation of the Left-Right positions of users in our panel in the recommender, which is also independent of other common attributes, such as age and gender. We explore the consequences of our findings by modifying these ideological representations in the recommender and analyzing the trade-off in resulting recommendations in terms of political leaning, diversity, and relevance of offered contents.