Biased training data produce results that confirms the bias. These end up being new data points. Following models trained in the same fashion will be even more biased.