[Paper] Measuring and Fostering Peace through Machine Learning and Artificial Intelligence

Source: arXiv - 2601.05232v1

Overview

The paper explores how machine‑learning (ML) and artificial‑intelligence (AI) can both measure the “peacefulness” of news and social‑media content and encourage more constructive media consumption. By turning raw text and video transcripts into quantitative peace scores, the authors build tools that let everyday users see how the media they read or watch may be inflaming conflict or fostering calm.

Key Contributions

• Cross‑domain peace detection: Neural‑network models that infer peace levels from news‑article embeddings and generalize across distinct news corpora.
• Social‑media peace metrics: Word‑level (GoEmotions) and context‑level (large language model) classifiers that assess peace‑related social dimensions in YouTube videos.
• Real‑time user feedback tool: A Chrome extension, MirrorMirror, that displays a live “peace score” while users watch YouTube videos, nudging them toward calmer content.
• Open‑source vision: The authors propose an extensible platform for journalists, creators, researchers, and platforms to audit and improve the tone of their media pipelines.

Methodology

1. Data Collection

   • News: Large sets of online articles labeled for peace‑related language (e.g., conflict‑free vs. conflict‑laden).
   • YouTube: Video transcripts paired with human‑annotated emotion tags (using the GoEmotions taxonomy) and broader social‑dimension labels (e.g., respect, cooperation).

2. Feature Extraction

   • Text embeddings from pre‑trained transformer models (e.g., BERT) capture semantic nuance in news articles.
   • For videos, two parallel streams:
     • Word‑level: Count‑based emotion features from GoEmotions.
     • Context‑level: Prompted large language models (LLMs) that generate a peace probability given the full transcript.

3. Model Training & Validation

   • Supervised classifiers (feed‑forward neural nets) trained on the news dataset; performance tested on a separate news source to verify transferability.
   • Multi‑task learning for YouTube data, jointly predicting emotion intensity and peace score.

4. MirrorMirror Extension

   • The trained YouTube model runs client‑side in the browser, scoring each video in real time.
   • UI overlays a simple gauge (e.g., green‑yellow‑red) and optional textual explanation of why a video is deemed “less peaceful.”

Results & Findings

These numbers demonstrate that automated peace scoring is both reliable across media types and actionable for end users.

Practical Implications

• For developers: The open‑source models and the MirrorMirror codebase can be integrated into any browser‑based media platform, enabling real‑time tone dashboards without heavy server infrastructure.
• For content creators: Peace scores act as a feedback loop, encouraging creators to balance emotional hooks with constructive messaging—potentially improving audience trust and platform reputation.
• For newsrooms: Automated peace metrics can flag articles that may unintentionally amplify conflict, supporting editorial checks before publishing.
• For platforms: Peace‑aware recommendation algorithms could diversify feeds, reducing echo‑chambers driven by outrage‑based content.
• For researchers & NGOs: Quantitative peace indices open new avenues for large‑scale studies of media influence on social cohesion and conflict dynamics.

Limitations & Future Work

• Cultural bias: The training data are predominantly English‑language and Western‑centric; peace perception may differ across cultures.
• Granularity: Current scores are coarse (peace vs. non‑peace) and may miss nuanced sub‑dimensions such as “constructive criticism.”
• User adoption: MirrorMirror’s impact hinges on users actively installing and trusting the extension; broader platform integration is needed.

Future directions

• Expand multilingual datasets and evaluate cross‑cultural validity.
• Refine the taxonomy to include more fine‑grained social dimensions (e.g., empathy, solidarity).
• Explore API‑based services for seamless embedding into news CMSs and video platforms.

Authors

• P. Gilda
• P. Dungarwal
• A. Thongkham
• E. T. Ajayi
• S. Choudhary
• T. M. Terol
• C. Lam
• J. P. Araujo
• M. McFadyen-Mungalln
• L. S. Liebovitch
• P. T. Coleman
• H. West
• K. Sieck
• S. Carter

Paper Information

• arXiv ID: 2601.05232v1
• Categories: cs.CL, cs.CY, cs.LG
• Published: January 8, 2026
• PDF: Download PDF