It can helps to classify any potential hate speech (toxic) comments in notes (kind 1). It will gives prediction probability which can be used for filtering. Users can freely set their probabilty thresold (in percentage) as the filter settings on relay.

I have already implemented similar classification features in nostr-filter-relay which focus as "global feed" relay.