Source
Chuai & Zhao (2020), LIT-379 (arXiv 2004.10399v3), read in NOTE-852: Fig. 1 and Tables S2–S14 (distributions), Table 1 and Table S17 (regressions), S10 and Table S18 (English data), S13 and Table S26 (survey), S14 and Table S27 (the 20% threshold).
What was actually shown
The anger and joy gaps. On 22,479 fake posts (tagged by Weibo’s fact-checking committee) and 10,000 true posts (from verified accounts), scored with a hand-labelled lexicon of 6,155 terms, mean anger share is 0.165 in fake and 0.112 in true news, and joy 0.443 against 0.609 (Table S4). The gaps are larger between highly retweeted fake posts by authors with few followers (LHF) and lowly retweeted true posts by authors with many (HLT): anger 0.267 against 0.111, joy 0.329 against 0.611 (Table S2). Highly retweeted (≥ 10 retweets) posts carry more anger than lowly retweeted ones regardless of label (Table S5). Two classifiers and a rank measure over six more give the same direction for anger and joy (S7); disgust, sadness and fear do not behave consistently.
The association with retweets. In linear models of retweets within 48 hours, with follower, friend, mention, hashtag, location, date, URL, length, one emergency event and seven topics as controls, anger share has coefficient 23.96 and joy share −29.56 among fake posts (Table 1 (3), N = 20,323, R² = 0.084), and 22.28 and −35.98 among all posts (Table 1 (4), R² = 0.134). The emotion shares sum to 1, so these are relative to the omitted disgust, sadness and fear. Moving 0.1 of a post’s emotional vocabulary from joy to anger goes with about 5.4 more retweets among fake posts and 5.8 among all posts; the paper reports 5.8 for fake news, which is the all-news figure.
What could have come out the other way. The gaps could have vanished under other scoring methods (they did for the other three emotions), and the anger coefficient could have lost sign or significance once author reach and topic were controlled. Neither happened. On unlabelled English tweets anger also predicts reaching 10 retweets (logit 0.215, s.e. 0.062, Table S18).
What this does not say
- That anger causes fake news to spread. The paper says it does, “causally”, throughout. The evidence is regression on observational data with no experiment, instrument or natural experiment, and the models leave nearly 90% of the variance unexplained. Content that carries anger words may carry other features (alarm, a call to repost, a rumour type) that do the work.
- That fake news is angrier because it is false. Fake posts were labelled because they had spread widely, and true posts come from verified authors, so the fake-against-true comparison mixes falsity with selection on spread and on source.
- That the mechanism is platform-independent. The English evidence has no fake/true labels in the main test, a gap of 0.123 against 0.110 in a second set, and 478 tweets in a third.
- That anger works through anxiety management. The survey result rests on two anger-dominated posts, with unadjusted K-S tests at p ≈ 0.02–0.04, and measures stated motives, not reposting.
- That flagging angry posts would work. The 20% threshold is chosen in-sample, and its effect (46% of popular fake news caught, 22% of popular true news) assumes that no flagged post is ever retweeted.
Connections
- The emotion accounts in the record (THEORY-055, THEORY-057) concern what emotions are; this one concerns a word-level proxy for emotion in text, and depends on none of them.
- THEORY-123 holds that displayed mood spreads within small groups, shown experimentally. This account is about content features predicting sharing at scale, not about one person’s emotion passing to another.
- LIT-083 covers the response side of disinformation, this the spread.
Note of 2026-10-30: narrowed and promoted
Promoted as a report of what the corpus measured (ADR-065). The paper’s thesis, that anger makes fake news viral, is THEORY-1221, Proposed, which extends this entry and keeps its old condition.