One of the key points of that paper is that the body of written works is biased and those biases will be amplified by compressing that body into LLMs, with outlier data, with low coincidence rate, being suppressed as unreliable - data that is structurally dissimilar to the bulk is noise. Similarly to how PageRank suppressed nodes with low number of edges, probably contributing significantly to the homogenous corporate mall-internet we enjoy today.