I received this response after inquiring to 10x directly (summarized):

Other mechanisms could also contribute to intronic reads, such as internal poly-A priming (see the detailed technical note). By including intronic read counting, it will potentially have a bias for poly-A-rich genes (not referring to exons here). Aligning with introns increases the sensitivity of gene detection, since it helps the detection of genes that have intronic UMIs but 0 exonic UMIs in any cell.

Different samples have different intronic read mapping profiles (table 1 in above link). (Also, if there is a polyA stretch within an exon, then multiple 'unique' reads could potentially be generated from a single transcript). Therefore, if you expect a high fraction of intronic reads in your single-cell sample and would like to increase the gene detection sensitivity, you could include --include-introns.