A Large Scale Study of Misinformation on WhatsApp

Kiran Garimella

In brief

This talk presents a large-scale empirical study of misinformation spread on WhatsApp in India, Brazil, and Indonesia, covering 5 million messages across political groups. The work identifies three main categories of image-based misinformation (out-of-context images, false quotes, and photoshops), showing 13% of shared images are misinformation. The speaker discusses four approaches to mitigation: on-device fact-checking, machine learning models, community-driven fact-checking, and forwarding limits, each with distinct tradeoffs.

Key findings

Scale of problem: 13% of images shared in public political WhatsApp groups contain misinformation, with over 50% of WhatsApp content being non-textual (images/video).
Image misinformation taxonomy: ~70% of misinformation falls into three categories: old images taken out of context ("cheap fakes"), false quotes or statistics overlaid on images, and photoshopped fake news screenshots.
Persistence: Misinformation images last much longer than other content; 50% persist for over 100 days due to resharing.
Contextual spikes: Peak misinformation during geopolitical crises (India-Pakistan conflict) rather than elections; deepfake videos successfully distributed across 5,800+ groups reaching ~15M people during Delhi elections.
Cross-platform coordination: Political parties in India coordinate WhatsApp groups to drive hashtag trends on Twitter via shared Google Docs with template tweets.

Data & methodology

Collection: Automated tools deployed on Android phones to join and monitor public WhatsApp groups across three countries during election periods. Final dataset: 5 million messages, 10+ terabytes, ~1.5 million images.

Annotation: Three professional journalists manually labeled 2,500 unique images (clustered via perceptual hashing), finding ~800 flagged as misinformation. Inverse probability weighting applied to estimate prevalence across full dataset.

Solutions explored

On-device fact-checking: Hash-based matching of previously fact-checked images; 40% of Brazil's and 80% of India's misinformation shares could have been prevented, but not scalable and vulnerable to evasion.
Machine learning: Metadata-based features (who posted, where, how long content persisted) achieve ~0.67–0.78 accuracy; content features (object detection, OCR) fail across time.
Community-driven fact-checking: Leveraging in-group corrections observed in the dataset; live but risky if incentivized without safeguards.
Forwarding limits: WhatsApp's approach of restricting forwarding reduces dissemination velocity by ~10x but doesn't prevent determined sharing.

Notes

The work is valuable for its large-scale empirical grounding in an understudied platform and Global South context. The three-category taxonomy of image misinformation is robust (replicated in US 2018 midterms and UK during COVID-19). The mitigation solutions are pragmatic but each has limitations; the talk's honesty about tradeoffs is a strength. The cross-platform coordination finding opens a new research direction on information flows between encrypted and public platforms.

One limitation: the study focuses on public political groups, which may not be representative of private group dynamics where most WhatsApp activity occurs. The year/date of the talk itself is indeterminate from the transcript; context clues suggest it was delivered 1–2 years after events described (Brazil 2018 election, India 2019 elections, Delhi 2020 elections).