Panel: From Spam to Speech: Policing the Next Generation of "Unwanted Traffic"
summary
Automated moderation struggles with scale and nuance: off‑the‑shelf NLP averages 56-58% accuracy while tuned models rise to 75-80%, so false positives/negatives are common; hash matching handles limited copyright sets but fails at web scale; a vendor claims 99% detection for a narrow class of ISIS videos, not general propaganda. Machine learning generates convincing short texts and fake reviews that pass human judgment, while a prototype lets trusted friends moderate an inbox — useful in practice but not a substitute for broader policy or real‑time guarantees.
AI-written summary · about
chapters (9)
- 0:00Content moderation challenges
- 4:29Platform provider challenges
- 7:38Policy motivations
- 12:57User experience & harassment
- 18:35ML for content generation
- 22:51Automated detection state
- 32:25Root causes of harm
- 34:52Legal risks & collaboration
- 43:21Future of moderation
related talks
48mAI, truth, and the new information environment [Keynote]
36mCan AI Help Moderate In-game Chat?
16mFighting Harmful Content with AI: Real-World Safety
37mThe Shadows That Follow AI Generative Models
57mUsing AI to automate content moderation and compliance with 
50mPanel: Humanity in the Loop: Balancing AI, Data Science, and Human Ingenuity in the Cybersecurity..
06m7 practical tips for safer AI and language models
44mAdvancing Cybersecurity with AI and Machine Learning: Opportunities and Challenges