Short is the Road that Leads from Fear to Hate: Fear Speech in Indian WhatsApp Groups

Punyajoy Saha; Binny Mathew; Kiran Garimella

In brief

The speakers introduce fear speech — a distinct category of problematic speech that instills fear of a target group through harmful portrayals of their history, traditions, or future intentions — and present the first large-scale dataset and detection study of fear speech in Indian WhatsApp groups. They show that fear speech spreads more widely than regular hate speech but is less toxic and harder to detect, and that users exposed to fear speech are significantly more likely to believe anti-Muslim attitudes, suggesting fear speech serves as a pathway to radicalization.

Overview

Fear speech is an important but understudied category of problematic speech distinct from explicit hate speech. While hate speech directly expresses hostility and calls for violence, fear speech instills anxiety about a target group (e.g., "Muslims will dominate India") through subtle, indirect messaging. This subtlety makes fear speech harder to detect and moderate, yet its real-world consequences are severe.

Dataset and methodology

The speakers collected 1.4 million messages from 5,000 Indian political WhatsApp groups from August 2018 to August 2019 using a query-based search method. After filtering spam and deduplication, they annotated 1,000 fear speech messages and 3,600 non-fear speech messages using student annotators trained on detailed annotation guidelines. Inter-rater agreement (Fleiss' kappa) was 0.36 (fair agreement), reflecting the difficulty of the task. The dataset captures public political WhatsApp conversations in Hindi and English during the lead-up to India's 2019 elections.

Key findings

Characterization: Fear speech spreads to significantly more users and groups than non-fear speech and has a longer message lifetime. Emojis play a prominent role, with communal symbols (Hindu religious icons, Islamic crescent) appearing in fear speech co-occurrence networks.

Topic modeling: LDA identifies ten major topics in fear speech, including conspiracy theories (e.g., "love jihad" — interfaith marriage narratives), claims of demographic threat ("Muslim population increasing at alarming rates"), and historical grievances ("Muslims responsible for communal riots").

User susceptibility: In a survey of ~22,000 users (50% with active Facebook accounts), users who post fear speech or are in fear-speech-sharing groups report significantly higher belief in anti-Muslim statements compared to control users, suggesting fear speech exposure drives attitude change.

Detection: Transformer-based models achieve AUROC 0.83 but precision of only ~0.51, indicating state-of-the-art models are not accurate enough for platform-scale deployment.

Notes

This work is one of the first to systematize fear speech as a distinct concept and scale, providing both a reusable dataset and detection baselines. The gap between AUROC (0.83) and precision (0.51) highlights a key challenge: even well-calibrated models fail at production-ready performance on subtle speech categories. The survey evidence linking exposure to belief change is valuable for understanding the radicalization pathway. The work was presented at VideoLecturesChannel and aligns with the "100 Video Talks on Misinformation" curated list.