Sift carries a short built-in list of offensive terms. It catches the common ones, plus their obvious variants, instantly. It's separate from the smart check's hateful-content read, which handles context and sneakier wording.
What it is for
Getting the worst words down right away instead of leaving them up while a model request runs. A few edge-case dodges (lookalike alphabets, zero-width characters, underscores) can slip past the list and land on the smart check instead.
What it is not
Your whole policy on hateful content. Borderline attacks and coded language still go through the smart check and your thresholds.