Let me try to clear your confusion: an input my seem innocent (e.g. zip code), but a zip code is likely to correlate to ethnicity and race in some regions. So even if the inputs seem legal, an ML model that’s sophisticated enough can derive illegal results that discriminate against certain populations.