Human knowledge is accumulated in several ways: through patents, scientific publications, encyclopedias, news, etc. In each case the involved “knowledge items” form a directed network that shows which item is built on which others. For example, patents (nodes) cite (link to) other patents (nodes). The usefulness of knowledge is most often measured on single knowledge items by article-level metrics (ALMs). In science the most common ALM is the citation number, n, quantifying impact. Instead of the impact here we discuss originality. We compute the probability, p, of directed links pointing from a node’s in-neighbors to its out-neighbors. Low values of p mean high originality. For several large real knowledge networks we find a very low correlation between n and p. Thus, we suggest that p provides qualitatively novel information about single knowledge items of human knowledge, such as patents, scientific publications, encyclopedia and news articles, etc.