Reranking partisan animosity in algorithmic social media feeds alters affective polarization

Today, social media platforms hold the sole power to study the effects of feed ranking algorithms. We developed a platform-independent method that reranks participants' feeds in real-time and used this method to conduct a preregistered 10-day field experiment with 1,256 participants on X during the 2024 U.S. presidential campaign. Our experiment used a large language model to rerank posts that expressed anti-democratic attitudes and partisan animosity (AAPA). Decreasing or increasing AAPA exposure shifted out-party partisan animosity by two points on a 100-point feeling thermometer, with no detectable differences across party lines, providing causal evidence that exposure to AAPA content alters affective polarization. This work establishes a method to study feed algorithms without requiring platform cooperation, enabling independent evaluation of ranking interventions in naturalistic settings.
, The dataset is collected with custom instrumentation through a browser extension, web surveys, and with in-feed surveys added to the participants' feeds on X.
, , # Reranking partisan animosity in algorithmic social media feeds alters affective polarization

## Contact

If you have any questions, please feel free to reach out to:

* Martin Saveski ([[e-mail]](mailto:[e-mail]))
* Tiziano Piccardi ([[e-mail]](mailto:[e-mail]))

## Setup

Before running the scripts:

1. Set the working directory to the code folder: `setwd(\"/path-to-repository/code/\")`
2. Configure the paths in `_constants.R`
3. Make sure that you have installed all the packages listed in `requirements.md`

## Scripts summary

* `01_attrition.R`: Attrition analysis
* `02_balance.R`: Covariate balance analysis
* `03_polarizaton_main.R`: Affective polarization analysis
* `04_emotions_main.R`: Emotions analysis
* `05_att_main.R`: Political attitudes analysis
* `06_polarization_hte.R`: Affective polarization, heterogeneous treatment effects analysis
* `07_engagement_analysis.R`: Engagement analysis
* `08a_pate_raking.R`: Population Average Treatment Effects ..., We take the original ID, concatenate it with a secret salt string, and hash the resulting string. Hashing ensures that the original IDs can’t be easily recovered, and adding the salt protects against dictionary-based attacks, where an attacker may have a list of Bovitz or CloudResearch IDs to hash and compare against the anonymized ones. We received user consent to publish the data in de-identified form.