Science1 publisher2 min readPublished
Reposts and replies peak at a moral-language density near 0.30 across three platforms
Researchers in Chile and the US found engagement across 1.62 million Twitter, Reddit and 8chan observations peaked at a moral-density score near 0.30. The data are observational, so they locate the peak without telling a writer whether trimming moral words from a post would earn more.
The Scientist · Science desk
What happened
- The team scored every post twice: moral loading, how much moral content it held, and moral density, how concentrated that content was across its words.
- Higher moral loading was consistently linked with more engagement on all three platforms.
- With moral loading held constant, higher moral density went with lower engagement.
- The work, by Universidad del Desarrollo, the University of Massachusetts Amherst and Northwestern University, appeared in Nature Human Behaviour.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- decision Framing a moral appeal splits into two choices, what concern to raise and how tightly to word it, and in this data only the tight wording went with less engagement.
- constraint Because 0.30 is an average across platforms and topics and the peak shifts with context, no single account or topic gets a threshold it can write toward.
- precedent Analyses that score moral language as a single count have reason to split it, since moral content and its concentration moved with engagement in opposite directions here.
Cristian Candia, the paper's first author, put the question narrowly. "Earlier research showed that moral language was associated with greater sharing online," he told Phys.org. "We wanted to understand whether that relationship had a limit. A message can express an important moral concern, but does filling it with moral language necessarily make it more engaging?" [9]
To answer it, the team had to separate a message's overall moral relevance from how saturated it is with moral language [15]. They used Distributed Dictionary Representations, a method that measures how closely text relates to the concepts in a predefined dictionary [10]. A pretrained model supplied the raw material. It represented words as numerical vectors based on their meanings and the contexts they appeared in [11]. Candia described the rest: "Using computational language analysis, we measured how closely messages related to moral concepts and how concentrated that language was. We then examined their relationship with retweets or replies, depending on the platform, accounting for available features such as links and multimedia. We also checked the findings using conventional moral-word dictionaries." [12]
Holding loading fixed is the design choice to notice. It lets the density score answer Candia's question directly: whether packing more moral language into a message adds engagement once the moral concern itself is accounted for [c7, c9]. The loading result agrees with the earlier work Candia cited, and the density result is the new part [17]. The peak comes from putting both in one model. "Our model brings these relationships together to predict an intermediate range of moral expression associated with the highest engagement, which varies across contexts," Candia said [13].
The roughly 1.62 million observations are not all the same kind of unit [2]. "The final analysis included individual posts on Twitter and Reddit and discussions grouped by topic and day on 8chan," Candia said [8]. The Twitter posts and their engagement figures came from previously released datasets. The Reddit and 8chan material came from archived public discussions [3]. The 0.30 figure averages over all of it, so a single tweet and a day of 8chan discussion on one topic both feed the same number [c4, c8].
The outcome was retweets or replies, depending on platform [12]. That counts people who acted on a post. Views, the figure closer to reach, are a different measurement.
The thing this doesn't tell you is how big the effect is. The Phys.org account gives the location of the peak but not how steeply engagement falls beyond it [4].
What to watch
- Effect sizes in the full Nature Human Behaviour paper, showing how far reposts and replies fall per step of density above 0.30.
- An experiment showing people one moral argument written at two densities would test whether the wording itself causes the drop.
- Whether the same peak appears in view counts and on current feeds, given that the Twitter data came from previously released datasets.