This category reads for hateful and discriminatory language in context, not just words out of a dictionary. It works alongside the instant offensive-terms list.
What it catches
- Dehumanizing language aimed at protected groups.
- Clear contempt tied to who someone is.
- Some coded or indirect wording, when the model's read is confident.
Intent softening
When a message is clearly a joke, a meme, or someone quoting another person, Sift can ease off one step for this category and a few like it. Block becomes review, review becomes pass. A serious attack on people doesn't get a pass because someone tacked lol on the end.