ML exists in a murky grey area between variance and bias. This has profound effects on how it handles classifying people. If the variance is too high, it is o

ML exists in a murky grey area between variance and bias. This has profound effects on how it handles classifying people. If the variance is too high, it is overfitting on individual characteristics, potentially leaking PII. If the bias is too high, it is likely inappropriately lumping an individual in with a disadvantageous crowd/outcome. There's nothing "intrinsic" about bias, and there's no way to effectively measure it universally. Simply handling African American facial features in a GAN would satisfy the situation in TFA, but there's undoubtedly another related bias issue lurking in the model. Inevitably, handling bias in ML winds up being akin to playing whack-a-mole, handling only the situations where people are noisily complaining about something. This has always bothered me, but I've never found a satisfactory approach for dealing with it.