Journeys through the content moderation landscape of video
summary
A practical runbook for video moderation: an open source upload flow (stream.new) sends five sampled images to Hive for NSFW/violence scoring and automatically deletes flagged uploads, which proves “good enough” for strict denial policies; live streams use the first ten minutes of frames with Google Vision to tag sports objects but that yields noisy false positives; clipping ten minutes of audio and running Whisper + ChatGPT identifies sport type reliably; outreach to uploaders sometimes stops piracy immediately.
AI-written summary · about
chapters (9)
- 0:00Content moderation intro
- 1:03What is moderation and why it matters
- 2:27Risk management in moderation
- 3:10Moderation for user uploads
- 5:14Live streaming piracy challenges
- 7:17Technical hurdles in live detection
- 9:00Audio analysis for live moderation
- 10:24Demo: Whisper audio detection
- 11:04Simple solutions & closing
related talks
52mLeading the way towards implicitly safe communities
39mThe True Cost of Unmoderated Collaboration: A Story from the Trenches
31mThe Virtual Is Real: Ending Online Violence
06mBeyond the West: Exploring the Impact and Regulation of Non-Consensual Image-Disclosure
33mBuilding Compassionate Communities in Tech
09mBecoming Feerless: Lessons Learned from Building Tools for Mental Health
23mTackling New Challenges in a Virtual Focused Community
30mOccupy Ruby: Why We Need to Moderate the 1%