Does social media exacerbate political polarization? Experimental evidence from varying access to Twitter/X during a presidential debate

We study how variation in social media access (Twitter/X) during a salient political event affects political polarization and stress, and how these exposure effects relate to broader patterns of selection into social media. We conducted a pre-registered laboratory experiment during the 2019 Argentine presidential debate, recruiting both Twitter users and non-users. Twitter users were randomly assigned to one of four conditions: watching the debate with their usual phone access, watching the debate with their phones switched off, watching the debate while interacting with an experimenter-run Twitter account that shared a balanced mix of partisan content, or watching a politically neutral documentary without phones. Non-Twitter participants were randomly assigned to watch the debate or the documentary without phones. Selection comparisons indicate that recruited Twitter users are more polarized than recruited non-users. Causal exposure comparisons within the Twitter-user sample show that two strategies often proposed to reduce political polarization on social media, removing access to the platform and exposing users to counter-attitudinal content, do not reduce polarization and may instead produce mild increases. The latter treatment also significantly increases physiological stress (as measured by salivary cortisol). Further exploratory analyses reveal substantial heterogeneity tied to users’ online political environments. Politically segregated individuals — that is, users embedded in echo chambers — exhibit clear backfire effects when access is removed or cross-ideological interactions are encouraged. In particular, both treatments increase polarization among these users. In addition, when cross-ideological interaction is encouraged, segregated users concentrate a larger share of their Twitter engagement within like-minded networks and show higher cortisol levels. By contrast, non-segregated users show little to no response and closely resemble non-users. Moreover, the selection differences observed between recruited Twitter users and non-users in our sample are concentrated among politically segregated users.