Fighting Harmful Content with AI: Real-World Safety
summary
Gcorp builds a moderation pipeline that detects CSAM and high-risk abuse at scale, scanning against a backdrop of class imbalance and ~500 hours of video per minute uploads. The stack uses frame CNNs, Whisper-style transcription, JU and 3D CNNs, containerized inference, and a Celery-like task queue called salary to priority-schedule live vs batch jobs. Traces record which model contributed, bias issues lead to retraining, and edge-optimized models target deployments under 200 mgabytes.
AI-written summary · about
chapters (11)
- 0:00Introduction and Speaker Background
- 0:39The Importance of AI in Content Moderation
- 1:36Challenges in Detecting Harmful Content
- 3:25System Architecture Overview
- 3:50Detailed Pipeline Breakdown
- 8:06Model Specifics and Their Roles
- 9:38Real-World Applications and Integrations
- 10:59Ethical Considerations and Privacy
- 12:59Future Developments and Innovations
- 14:57Key Takeaways and Lessons Learned
- 15:56Conclusion and Final Thoughts
related talks
57mUsing AI to automate content moderation and compliance with 
45mPanel: From Spam to Speech: Policing the Next Generation of "Unwanted Traffic"
36mCan AI Help Moderate In-game Chat?
31mSecurity Considerations for Services Using AI Models
48mStay Ahead of Adversarial AI in OT/ICS Environments – Mitigating CWE-1039
37mThe Shadows That Follow AI Generative Models
32mEmotional Trauma and Machine Learning
48mBuilding and using AI models responsibly