What sort of bias is being discussed here? Is it that humans might tend to highlight aspects which is informed by their own inclinations? Wouldn’t the same be true for the summarising model as well, because its knowledge is contingent on the data used for training (which is again created by humans)