A common pitfall in AI performance is " aggregate accuracy, " which masks poor performance in specific demographics or subgroups. To detect these " subgroup regressions, " auditors should look for " comparative evaluations " where the model is tested separately for different languages. " Parity thresholds " establish the maximum allowable difference in performance between the majority group (English) and minority groups. This ensures fairness and consistent user experience. Post-go-live human review is reactive, and translating everything to English ignores the nuances of the original language that likely caused the errors in the first place.