SOAP is a techno-legal framework for auditing systemic risks in Very Large Online Platforms (VLOPs) through sock-puppet auditing, designed to operate within the Digital Services Act (DSA) regulatory framework. The system enables researchers, regulators, and civil society actors to investigate how personalized recommender systems contribute to filter bubbles, political radicalization, and harmful content amplification.
This repository contains the source code for the SOAP system, as described in this paper:
From Walls to Windows: Creating Transparency to Understand Filter Bubbles in Social Media
Proceedings of the Second Workshop on the Normative Design and Evaluation of Recommender Systems (NORMalize 2024), co-located with the 18th ACM Conference on Recommender Systems (RecSys 2024), Bari, Italy, October 18, 2024.
- DSA-Compliant Data Collection: Scrapes only publicly available data accessible without user credentials, ensuring compliance with privacy expectations and regulatory frameworks
- Active Sock-Puppet Auditing: Employs configurable sock-puppet accounts that dynamically adapt behavior based on "primer prompts" to simulate diverse user personas and test recommendation system responses
- Multimodal Content Analysis: Integrates large language models for automated deductive coding across text, image, and audio content, reducing manual annotation burden while maintaining oversight
- Longitudinal Filter Bubble Tracking: Monitors the evolution of personalized feeds over time to enable causal experimentation and measure convergence toward content homogeneity
- Scalable Systemic Risk Investigation: Provides structured, repeatable audits of platform behavior under controlled conditions for empirical evidence of recommendation engine patterns
SOAP is currently under development, with new features and improvements being added. The following features are in progress:
- Adding new open-source multimodal large language models
- Expanding documentation
- Providing a guide for deploying SOAP on Cloud services and via cron jobs
SOAP operates within the DSA regulatory framework by accessing only publicly available data from platform Explore pages. The system currently runs on Google Cloud Vertex AI, utilizing the multimodal model Gemini 1.5 flash for automated content analysis. To run the code as-is, a Google Cloud account is required. The deductive coding component can be replaced with other models (either self-hosted or on another platform) to suit different research needs.
The following are needed to run the code as-is:
- Sock-puppet account credentials (username and password). It is recommended to use dedicated research accounts as there is a risk of accounts being flagged for automated behavior, potentially resulting in restrictions or bans.
- Google Cloud:
GOOGLE_APPLICATION_CREDENTIALSBUCKET_NAMEPROJECT_NAME- API credits for both storage (bucket) and LLM usage
Legal and Ethical Considerations: SOAP is designed with legal compliance in mind, accessing only publicly available data and employing passive interaction patterns (viewing, liking) without active discourse (commenting, sharing, messaging). The system includes safeguards against harmful prompting and is intended solely for public interest research purposes.
The system is orchestrated and managed by the orchestrator.py script, which coordinates all components for sock-puppet auditing as shown in "Figure 1: Agent Scenario for Creating Filter Bubbles based on Primer Prompts." The framework employs active sock-puppets that dynamically adapt their behavior based on primer prompts to test how recommendation algorithms respond to different user interest profiles.
The following table outlines the key classes and their respective functions:
| Class | Function |
|---|---|
explorefeed_scraper.py |
Scrapes Explore feed data from Instagram using DSA-compliant methods |
saveFeedCloud.py |
Saves scraped feed data to Google Cloud Storage |
VertexAi.py |
Utilizes Google Cloud Vertex AI for multimodal content analysis and automated deductive coding |
steering_wheel.py |
Controls sock-puppet interactions (liking, reporting, archiving) to simulate user behavior patterns |
SOAP enables investigation of various systemic risks including:
- Filter Bubble Formation: Tracking how quickly platforms converge toward content homogeneity
- Content Amplification Patterns: Measuring exposure to potentially harmful or misleading content
- Algorithmic Bias Detection: Testing recommendation system responses across different user personas
- Regulatory Compliance: Supporting DSA oversight by providing empirical evidence of platform behavior
Before deploying a new prompt for creating filter bubbles, it is recommended to conduct reliability testing with the prompt and the multimodal language model (LLM). Code for both inter-reliability and intra-reliability testing can be found in the /reliability-evaluation/ directory.
If you use/modify this source code or refer to our paper, please add a reference to our publication:
Luka Bekavac, Kimberly Garcia, Jannis Strecker, Simon Mayer, and Aurelia Tamò-Larrieux. 2024. From Walls to Windows: Creating Transparency to Understand Filter Bubbles in Social Media. In NORMalize 2024: The Second Workshop on the Normative Design and Evaluation of Recommender Systems, co-located with the ACM Conference on Recommender Systems 2024 (RecSys 2024), October 18, 2024. Bari, Italy. https://www.alexandria.unisg.ch/handle/20.500.14171/120987
@inproceedings{bekavac2024,
title = {From {{Walls}} to {{Windows}}: {{Creating Transparency}} to {{Understand Filter Bubbles}} in {{Social Media}}},
booktitle = {{{NORMalize}} 2024: {{The Second Workshop}} on the {{Normative Design}} and {{Evaluation}} of {{Recommender Systems}}, Co-Located with the {{ACM Conference}} on {{Recommender Systems}} 2024 ({{RecSys}} 2024)},
author = {Bekavac, Luka and Garcia, Kimberly and Strecker, Jannis and Mayer, Simon and {Tam{\`o}-Larrieux}, Aurelia},
year = {2024},
month = oct,
address = {Bari, Italy},
abstract = {Social media platforms play a significant role in shaping public opinion and societal norms. Understanding this influence requires examining the diversity of content that users are exposed to. However, studying filter bubbles in social media recommender systems has proven challenging, despite extensive research in this area. In this work, we introduce SOAP (System for Observing and Analyzing Posts), a novel system designed to collect and analyze very large online platforms (VLOPs) data to study filter bubbles at scale. Our methodology aligns with established definitions and frameworks, allowing us to comprehensively explore and log filter bubbles data. From an input prompt referring to a topic, our system is capable of creating and navigating filter bubbles using a multimodal LLM. We demonstrate SOAP by creating three distinct filter bubbles in the feed of social media users, revealing a significant decline in topic diversity as fast as in 60min of scrolling. Furthermore, we validate the LLM analysis of posts through an inter- and intra-reliability testing. Finally, we open source SOAP as a robust tool for facilitating further empirical studies on filter bubbles in social media.}
}If you have questions about the prototype or the publication, feel free to contact Luka Bekavac (lukajurelars.bekavac@student.unisg.ch).
This research was conducted by the Interaction- and Communication-based Systems Group (interactions.ics.unisg.ch) at the University of St.Gallen (ics.unisg.ch) in collaboration with the Legal Design & Code Lab (https://wp.unil.ch/legaldesignandcodelab/) at the University of Lausanne (https://unil.ch).
All source code in this repository is licensed under the Apache License 2.0 (see LICENSE) if not stated otherwise. Included third-party code may be licensed differently (see the respective files and folders).
