Nudge-based misinformation interventions are effective in information environments with low misinformation prevalence

The study investigates whether nudge-based interventions—specifically a combined accuracy prompt and social-norm message—effectively reduce engagement with misinformation in social-media environments that more closely reflect real-world conditions (where false content is a small fraction of overall content). The research addresses concerns that prior experiments often use unrealistically high misinformation proportions (e.g., 50%) and only verifiable news headlines, potentially misrepresenting nudge efficacy, which can be context-dependent and subject to decay. The authors pre-registered two hypotheses: (1) the nudge would improve engagement discernment (liking and sharing of true versus false headlines), and (2) nudge effectiveness would depend on misinformation prevalence, with a tentative expectation that it might be greater when misinformation is more prevalent.

Prior work shows several psychological interventions reduce misinformation belief and sharing, including debunking and generalized, scalable nudges that prompt accuracy consideration or leverage social norms. Accuracy nudges demonstrate small but replicable improvements in sharing discernment, though effects may be context-sensitive and potentially decay. Studies indicate information composition (true:false base rates) biases judgments and can moderate intervention efficacy. One recent study including social posts found modest improvements in sharing discernment from accuracy prompts but no reduction in false engagement when proportions were fixed, leaving open whether varying misinformation prevalence alters nudge impact. The limited-attention utility model suggests nudges increase attention to veracity, potentially improving decision quality; however, realistic social-media feeds contain low misinformation and large volumes of non-news content, making generalization from lab paradigms uncertain.

Design: Pre-registered 2×3 between-subjects design with factors nudge (present vs. absent) and misinformation proportion (50%, 20%, 12.5%). The task included both true and false headlines, making it a 2 (nudge) × 3 (misinformation proportion) × 2 (headline veracity) between-within design. Primary outcomes were sharing and liking discernment (differences in engagement with true vs. false headlines). Participants: 1501 U.S.-based Prolific participants with at least one social-media account were recruited; after pre-registered exclusions (effort, dwell time criteria, etc.), the final sample was N=1387 (M_age=41.39, SD=14.14; 736 females, 621 males, 26 non-binary, 2 transgender men, 1 transgender, 1 no gender). Condition ns: No nudge—12.5%:221; 20%:250; 50%:238. Nudge—12.5%:223; 20%:223; 50%:232. Materials: Nudge combined accuracy and social-norm prompt emphasizing the importance of sharing accurate content and that >80% of U.S. adults value sharing accurate information; followed by an explicit accuracy rating of a false headline on a 0–10 scale. Posts: Depending on condition, participants viewed (1) 50% misinformation: 40 true + 40 false news headlines; (2) 20% misinformation: 40 true + 10 false headlines; (3) 12.5% misinformation: 40 true + 10 false headlines + 30 social posts (non-falsifiable). Headlines were curated and piloted for believability, political leaning, shareability, and currentness; true headlines from reputable sources, false from non-reputable sources; images paired and sources fixed across participants. Social posts were author-generated and non-controversial. Simulated social-media feed: Posts displayed on a single scrollable page; participants could like, share, comment, or skip; follower counts and prior engagements were probabilistically assigned and updated. Procedure: After consent and instructions, participants (all told to engage as they would and try to maximize follower count) completed the feed task. Subsequently, they rated belief in a subset of headlines (maintaining condition-specific true:false proportions), reported political orientation, and were debriefed. Ethics: Approved by University of Western Australia HREC (2019/RA/4/20/6423). Analysis: Conducted in R (lme4 glmer); due to convergence issues with the pre-registered ordinal engagement models, sharing and liking were analyzed as separate binary outcomes via logistic mixed-effects models. Type-3 ANODEs (car::Anova) tested main effects and interactions; emmeans used for post-hoc contrasts. Random-effects structures were maximized as justified. Exploratory analyses assessed potential decay over post order.

• Sharing discernment: Significant nudge × headline veracity interaction (Type-3 ANODE: χ²=7.15, df=1, p=.007), indicating the nudge increased sharing of true relative to false headlines overall. No significant three-way interaction with misinformation proportion (χ²=1.51, df=2, p=.470), suggesting nudge efficacy did not differ significantly across proportions. There was a significant misinformation proportion × headline veracity interaction (χ²=8.96, df=2, p=.011).
• False vs. true sharing (collapsed across proportions): Nudge reduced sharing of false headlines (OR=1.44, SE=0.22, z=2.42, p=0.016, indicating fewer false shares in the nudge condition); no significant effect on sharing of true headlines (OR=0.95, SE=0.12, z=0.43, p=0.667).
• By condition: Nudge improved sharing discernment in 12.5% (χ²=7.04, p=0.008) and 20% (χ²=10.92, p=0.001) misinformation conditions; 50% condition showed a nonsignificant trend (χ²=3.49, p=0.062). Nudge marginally reduced false-headline sharing in 12.5% (OR=1.76, SE=0.49, z=2.05, p=0.040) and 20% (OR=1.70, SE=0.40, z=2.22, p=0.026) conditions; no effect on true sharing (ps>0.666).
• Baseline differences across proportions (collapsed over nudge): False sharing was lower in 12.5% vs. 20% (OR=0.48, SE=0.09, z=−3.92, p<.001); neither 12.5% nor 20% differed significantly from 50% for false sharing (ps>0.067). True sharing was lower in 12.5% vs. 20% (OR=0.54, SE=0.08, z=−4.11, p<.001) and vs. 50% (OR=0.56, SE=0.08, z=−3.83, p<.001); 20% vs. 50% did not differ (p=0.785).
• Liking discernment: No significant nudge × veracity interaction (χ²=3.50, df=1, p=.061) and no three-way interaction (χ²=0.07, df=2, p=.968). Significant main effects: headline veracity (χ²=22.15, df=1, p<.001; true > false) and misinformation proportion (χ²=12.60, df=2, p=.002). Overall liking was higher in 20% than 12.5% (OR=0.65, SE=0.08, z=−3.43, p=.002) and higher in 20% than 50% (OR=1.34, SE=0.17, z=2.37, p=.036), potentially driven by fewer total headlines in the 20% condition.
• Decay over time: Exploratory models including post order found no significant nudge × order or nudge × veracity × order interactions on sharing (ps ≥ 0.138), suggesting no strong evidence of linear decay across the session.

The nudge-based intervention improved sharing discernment primarily by reducing sharing of false headlines, with no significant effect on liking discernment. Contrary to expectations, nudge effectiveness was at least comparable—and in condition-specific analyses, statistically clearer—in environments with lower misinformation prevalence (12.5% and 20%) compared to a 50% misinformation environment. One potential explanation is that a high false-content base rate may implicitly prime accuracy considerations, diminishing the incremental impact of an explicit nudge in balanced (50%) settings. The results support the limited-attention utility model, implying inattentiveness to accuracy drives misinformation sharing and can be addressed by simple prompts. Engagement patterns across conditions also suggest participants prefer interacting with social (non-news) content, potentially reducing willingness to share news in mixed feeds. Overall, findings provide ecologically valid support for implementing scalable nudges on social media to improve the quality of information propagation, particularly via sharing behavior.

This study demonstrates that a combined accuracy and social-norm nudge can yield small but meaningful improvements in sharing discernment in simulated social-media environments, especially when misinformation is relatively rare—a common real-world scenario. The intervention primarily reduces sharing of false content without suppressing sharing of true content, indicating a beneficial shift in propagation behavior. Given the lack of evidence for rapid decay across the session and the alignment with the limited-attention utility model, large-scale deployment of such nudges on social platforms is supported. Future research should assess durability over longer timeframes or repeated exposures, generalize to diverse content types beyond headlines, examine how platform incentive structures affect nudge efficacy, and test performance in environments with even lower misinformation prevalence.

• Temporal proximity: The nudge was delivered shortly before the feed task, limiting inferences about long-term durability; exploratory order analyses were underpowered and showed no linear decay but numerical attenuation over time in some conditions.
• Stimuli scope: Effects were tested on news headlines; results may not generalize to other misinformation formats (e.g., ambiguous or subtle content).
• Incentive structures: Behavioral feedback (follower changes, prior engagements) was standardized across post types to avoid confounds; real-platform incentives may differentially affect behavior and nudge efficacy.
• Misinformation prevalence: Even the reduced proportions (12.5%, 20%) may exceed levels in some platforms; efficacy in environments with negligible misinformation remains to be fully assessed, though current results suggest nudges are unlikely to harm engagement quality.
• Model deviations: Pre-registered ordinal engagement models did not converge; analyses used binary outcomes for sharing and liking, which may impact comparability with some prior work.