Toxicity Prevention Study Platform

This tool is designed to study incivility in online debates and to test whether AI-based interventions can improve the quality of discussions in realistic social network environments. It allows researchers to examine how toxic or uncivil comments emerge and how they influence broader conversation dynamics.

In this case study, the platform was used to test an AI-assisted paraphrasing intervention. When users wrote potentially uncivil comments, they were provided with an alternative, AI-generated paraphrased version designed to reduce toxicity while preserving the original meaning. Users could then decide whether to post their original comment or accept the suggested revision.

The study evaluated whether these AI-supported interventions reduced toxicity at both the individual and conversation levels. Toxicity was measured using automated scoring (e.g., via Perspective API), capturing dimensions such as general toxicity, severe toxicity, insults, profanity, threats, and identity attacks. Results showed that paraphrased comments were significantly less toxic than the original versions.

Importantly, the tool also allows researchers to examine downstream effects. The findings indicate that when users accept AI-generated paraphrases, not only does their own comment become less toxic, but the overall tone of subsequent discussion improves as well. This makes the platform suitable for testing proactive, user-centered, and non-intrusive AI interventions aimed at fostering healthier online discourse in realistic social media settings.

Participant Flow

1. Landing page
2. Consent form
3. Pre-survey
4. Topic selection (choose 4 topics)
5. Username selection
6. Feed and per-post interaction loop:
   • You are shown one post/article at a time
   • You write a comment for that post
   • The system provides an LLM suggestion to reduce toxicity
   • You accept or reject the suggestion
   • You complete a questionnaire for that specific post
   • You are then moved to the next post
7. Repeat this loop until all 4 posts (from the 4 selected topics) are completed
8. After the 4th post questionnaire, you are redirected to the post-survey

App Structure

• client/: React frontend for participant flow pages and feed interactions
• server/: Node/Express backend, APIs, and MongoDB models
• server/routes/: API route definitions
• server/controllers/: request handlers and study logic
• server/models/: MongoDB schemas for users, posts, surveys, and interaction data
• screenshots/toxicity screenshots/: UI screenshots used in this README

Screenshots by Study Step

1) Landing Page

2) Consent Form

3) Pre-Survey

4) Topic Selection

5) Username Selection

6) Feed

7) Comment Editing with LLM Suggestion

8) Post/Article Questionnaire

9) Post-Survey

Local Development (Quick Start)

1. Ensure MongoDB is running and environment variables are configured.
2. Install server dependencies:
   • cd server && npm install
3. Install client dependencies:
   • cd client && npm install
4. Run server and client (in separate terminals):
   • Server: cd server && npm start
   • Client: cd client && npm run build

Deployment Note

The production deployment is available on:

• https://socialapp.ijs.si/

When started on your server, backend logs may show entries similar to:

• Server started on port 1077 and https://socialapp.ijs.si/
• Connected to the database successfully!

License

This project is licensed under the MIT License. See LICENSE.