Keeping the feed safe
A feed cannot just rank for engagement. It must keep harmful and policy violating content out, spam, abuse, and banned material. Moderation is a stage the pipeline cannot skip.
When moderation runs
- At ingestion, when a post is created, automated classifiers scan text, images, and video for clear violations and can block or hold the post before fan out.
- At serving, the feed pipeline filters out content later flagged or from blocked authors, so removed items vanish even after they were fanned out.
Running checks at both points catches violations that slip past the first gate.
How decisions are made
- Automated models score content for categories like spam and abuse, acting on high confidence and queuing the rest.
- A human review queue handles uncertain or appealed cases, and their labels train the models.
- User reports feed signals that re evaluate already published content.
The tradeoffs
Moderation balances speed, accuracy, and scale. Blocking too aggressively removes good content, blocking too little lets harm through. Most systems combine fast automated filters with slower human judgment.
Key idea
Feed moderation filters harmful content at ingestion and again at serving, combining automated classifiers, human review, and user reports to balance safety, speed, and scale.