How to fix policy violations while preserving content quantity and quality
summary
Discusses moderation strategies for user-generated images, comparing strict blocking to hiding violations and arguing that masking violations can improve user experience. Describes using blur as a moderation tool and reports reducing manual image reviews tenfold by applying blur-based masking techniques.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 2:58Manual moderation issues
- 4:22Choosing violations to fix
- 6:27Data collection and labeling
- 9:00Model training
- 11:10Inference optimization
- 13:21Moderation service architecture
- 15:50Production model optimization
- 19:22Classifier results
- 21:09Inpainting vs blur
- 22:11Inpainting model selection
- 27:12Speed and quality comparison
- 27:55Inpainting in production
- 30:02Inpainting results
- 31:07Issues and improvements
- 33:00Q&A
related talks
24mAlgorithms aren't neutral Making UX inclusive
28mIdentifying and Addressing Usability Vulnerabilities
01h 13m“Get out of my feed!”: Child-centered strategies to enhance well-being and agency on social media
25mAssume Worst Intent (Designing for the Abusive Ex)
26mDisruptive Design: Harmful Patterns and Bad Practice
01h 19mEffective Transparency and User Controls: Are We Close to a Breakthrough?
40mDesigning with Community Safety in Mind
57m(Ab)user Experience: The Dark side of Product and Security