Social highlighting study finds hidden factions among document readers

A new study by Kazuki Nakayashiki and Keisuke Watanabe, published on arXiv (2606.11613), investigates whether social highlighting—when many people highlight the same document—reflects a unified consensus or reveals hidden sub-groups. Using a margin-preserving curveball null model, they analyzed highlighting data from a co-readership platform. In Experiment 1, they found that within a single document, readers form strong sub-groups: pairs of readers agreed on highlights far beyond what shared salience, mark density, or sentence popularity would predict (nearest-neighbor agreement z=+6.3, significant in 88% of documents). When controlling for eight coarse document regions, shared engagement with those regions accounted for about 40% of the excess agreement. The remaining majority (z=+3.6, 77% significant) is finer, reader-specific agreement. This suggests the crowd within a document is factional—readers cluster into distinct groups based on what they highlight.

Experiment 2 tested whether these grouping patterns are a stable reader trait across documents. The cross-document split-half reproducibility of pair agreement was near zero overall (+0.078 and 0.000 in two samples). A power calibration showed the test is only informative for pairs that co-read many documents. In the small high-overlap subset (k>=4), point estimates were positive but imprecise, never significant, and weakened under the region-preserving null. The researchers honestly conclude that cross-document stability remains unresolved: the data is consistent with situational grouping or a weak-to-moderate stable trait. The crowd is factional within a document, but whether factions follow a reader across documents is beyond their current reach.