Main themes
Misinformation and counter-misinformation during COVID-19
During the early pandemic, Kumar's group collected 8 million COVID-19-related tweets to study misinformation dynamics. They identified ~33,000 misinformation tweets and a similar volume of counter-misinformation tweets from ordinary users. A key finding: 96% of counter-misinformation messages came from everyday users, not just professional fact-checkers—but the countering was often ineffective (rude, without evidence) and potentially counterproductive. They advocated for empowering users to counter misinformation with polite, evidence-based, empathetic responses instead.
Hate speech contagion and counter-speech
During COVID-19, hate speech targeting Asians surged online in correlation with real-world violence. By analyzing 227 million tweets, the group identified ~1.3 million hate speech tweets and ~1.1 million counter-speech tweets. They discovered hate is more contagious than expected—users with hateful neighbors are more likely to adopt hateful behavior. Counter-speech showed a modest negative effect on hate contagion (reducing it), providing early evidence that countering efforts can help, though more research is needed.
Graph-based detection with temporal dynamics
Most detection work focuses on content (text/NLP) alone. Kumar's group developed GEODE, a graph neural network that learns dynamic embedding trajectories for users and posts in temporal interaction networks. By encoding both network structure (who interacts with what) and temporal patterns (when interactions occur), GEODE outperforms content-only and even other graph-based methods by ~12% on Reddit and Wikipedia datasets for identifying malicious accounts and banned users.
Production impact and real-world deployment
- Flipkart (e-commerce fraud): Graph-based methods deployed to detect fake reviewers
- Wikipedia (ban evasion): Developing tools to help Wikimedia moderators identify evader accounts reusing network patterns
- Twitter Birdwatch: Twitter's community notes program was influenced by Kumar's findings that ordinary users form 96% of counter-misinformation efforts
Adversarial robustness
Deep learning detection models, including Facebook's proprietary system, are vulnerable to adversarial attacks. Kumar's group showed that even the best production models (e.g., Facebook's TIES system for malicious account detection, ~66% baseline accuracy) degrade significantly (~55% accuracy) under adversarial perturbations. This highlights a critical gap: detection systems must be not only accurate but also robust to manipulation by determined adversaries.
Key insights
-
Content alone is insufficient. NLP-based detection (embeddings, BERT) works, but fusing graph structure (user networks, interaction patterns) with content yields substantially better results.
-
Ordinary users matter. The majority of counter-misinformation comes from everyday users, not institutions. Platforms should design for and empower user-generated counter-speech.
-
Contagion is asymmetric. Hate spreads faster through networks than counter-speech does, but counter-speech still has measurable (if modest) protective effects.
-
The adversary adapts. Even production-deployed detection systems can be evaded. Research must continue to make systems more robust.
-
Multi-platform dynamics. Misinformation spreads across platforms (YouTube, Twitter, Facebook, Reddit). Single-platform analysis misses the full picture.
Notes
This talk bridges content analysis and network science, two previously siloed approaches in misinformation research. The emphasis on ordinary users as counter-misinformation actors is refreshing and aligns with platforms like Twitter's Birdwatch (now Community Notes), which systematize user-generated corrections. The adversarial robustness results are sobering: if deployed detection systems can be evaded, the field needs continuous improvement cycles and red-teaming cultures. The multi-platform observation is prescient—coordinated campaigns and information laundering across platforms remain understudied.