ML algorithms assume that the data on which they have been trained are drawn from some underlying distribution and that the unseen data on which they make predictions are also drawn from the same distribution. If an ML system encounters real-world data that falls afoul of this assumption, all bets are off as to the predictions.
