There's an acronym: OOD - out of distribution - for these situations.
There's no reason YT can't detect chess from hate speech if they updated their training set. Maybe they weren't aware of this failure case, or they didn't get to fix it, or by trying to fix it they caused more false negatives. The way they assign cost to a false positive vs a false negative is also related.
There's no reason YT can't detect chess from hate speech if they updated their training set. Maybe they weren't aware of this failure case, or they didn't get to fix it, or by trying to fix it they caused more false negatives. The way they assign cost to a false positive vs a false negative is also related.