Emotional language reduces belief in false claims

Samantha C. Phillips; Sze Yuh Nina Wang; Kathleen M. Carley; David G. Rand; Gordon Pennycook

doi:10.1017/jdm.2025.10019

In brief

Emotional appeals are assumed to increase belief in misinformation, but three experiments show the opposite: highly emotional false headlines are believed significantly less than low-emotion versions. Emotional language also reduces the perceived helpfulness of fact-checks. The results suggest emotionality signals unreliability rather than persuasiveness, with implications for fact-checking design and interventions.

Contributions

Demonstrates experimentally that emotional language decreases (not increases) belief in false claims when emotionality is varied while holding factual content constant
Shows fact-checking reduces the belief-reducing effect of emotionality for already-skeptical claims
Extends findings to real-world Twitter/X Community Notes data: emotionally-worded corrections are rated less helpful
Challenges the prevalent assumption that emotional language enhances misinformation effectiveness

Method

Three studies test emotionality and belief:

Studies 1a–1c (online experiments). Participants rated belief in claim pairs where emotionality was manipulated while core fact remained identical. Study 1a/1b used Prolific and Mechanical Turk (N=1005 and 987); Study 1c added a fact-check condition. Emotional language was induced via lexical and stylistic markers (e.g., "these greedy companies are destroying the planet" vs. "100 corporations are responsible for 71% of emissions"). Participants rated emotionality and accuracy on 7-point Likert scales. Analysis used linear mixed-effects models with random effects for participants and items.

Study 2 (observational). Analyzed publicly available X Community Notes (crowdsourced fact-checking, ~81k notes across ~45.6k tweets from 2021–2023). Extracted sentiment (positive/negative) using roBERTa-based model; outcome was helpfulness rating. Regression tested associations between emotional language and note ratings, controlling for tweet and note content features.

Results

Studies 1a–1c: High-emotionality versions of false claims elicited significantly lower belief than low-emotionality versions (Study 1a: β = -0.63, p < 0.001; Study 1b: β = -0.16, p < 0.001). The effect replicated across claim pairs. Participants rated high-emotion claims as less accurate. Fact-checks were more effective for already-skeptical claims; emotionality effect persisted for true claims (β = -0.14, p < 0.01), showing emotionality reduces discernment overall.

Study 2: Notes with more emotional language (both positive and negative sentiment) were rated significantly less helpful (positive emotion: exp(β) = 1.0052, CI [1.005, 1.0055], p < 0.001; negative: exp(β) = 1.0126, p < 0.001). Increased emotional language from 10th to 90th percentile decreased helpfulness probability from 0.63 to 0.49 (positive) and 0.43 to 0.29 (negative). More emotional tweets predicted higher engagement (retweets, likes), consistent with prior work showing emotionality drives sharing independent of accuracy.

Notes

Strong contribution that inverts a widely-held assumption: the paper carefully disentangles emotionality (via linguistic manipulation) from other factors that co-vary with emotional content, using both lab control and real-world observational validation. The fact that emotionality hurts even fact-check ratings suggests emotional tone signals low credibility to readers. Limitations: focus on negative/angry emotions; unclear whether results hold for discrete emotions (fear, moral outrage); field study limited to X. Design implications for fact-checking platforms are promising but would benefit from downstream behavioral tests (do emotionally-neutral corrections actually change minds better?).