True or false? Linguistic and demographic factors influence veracity judgment of COVID-19 rumors

Rumors—unverified information circulating under uncertainty—proliferated during the COVID-19 pandemic. Prior work has largely emphasized psychological and social determinants of rumor belief, transmission, and correction, with comparatively fewer studies probing how content characteristics of rumors and reader demographics affect processing and truth discernment. The present study asks two questions: (a) which content characteristics of COVID-19 rumors (e.g., length, presence of place/time/source cues, calls to action, veracity) influence reading times and veracity judgments, and (b) which participant demographic variables (gender, age, family socioeconomic status, and years of university study) impact veracity judgment. The authors hypothesized that both content features and demographic factors would shape processing (reading time, response time) and outcomes (accuracy) in a veracity judgment task. Understanding these factors is important for theory and for designing more effective rumor correction strategies.

The literature documents extensive interest in how rumors and misinformation spread, are perceived, detected, and countered. Psychological drivers of misinformation belief and resistance to correction, analytic thinking, endorsement cues, and social media context have been shown to influence truth discernment. Research on content features, though less common, indicates that properties such as length, presence of concrete places, dates/times, and calls to action can correlate with rumor veracity. Sentence processing research shows longer sentences increase reading time and that plausibility influences processing speed. Demographic factors such as age, education, and income have been associated with susceptibility to health misinformation in survey studies, with income (a component of SES) often linked to exposure and susceptibility. However, experimental evidence jointly examining content characteristics and demographics on rumor processing is scarce, motivating the present work.

Design and measures: Online experimental study measuring (1) reading times for rumor statements, (2) accuracy of veracity judgments (true vs. false), and (3) response times to the veracity probe. Participants: 121 undergraduate students from local universities in China were recruited via an online platform; 9 were excluded for low accuracy on catch trials (<50%), yielding N = 112 (11 male, 101 female). Mean age = 20 years (range 18–45). Years of university study: 1–4 (mean 1.7). All were native Chinese speakers (Mandarin learned <6 years old), with self-rated intermediate English proficiency. Family socioeconomic status (SES) was computed using the Hollingshead (2011) four-factor index, adapted by Kang et al. (2021); range 8–61 (mean 27). Ethical approval was obtained (CZLS2022222-A), and informed consent was provided; participants received compensation. Materials: 112 Chinese statements related to COVID-19 (102 experimental, 20 catch) sourced from the Chinese rumor-debunking platform piyao.org.cn. Headlines and visuals were removed; only the statement text was presented. Of the experimental items, half were true and half false as labeled by the platform. Content-characteristic predictors coded for analyses included veracity (true/false), length (number of Chinese characters), presence of concrete place, presence of specific time/date, presence of source cue, and presence of calls to action. Procedure: Data were collected in October–November 2022 using Gorilla (gorilla.sc). Trials were randomized. Each trial began with a 500 ms fixation, followed by the statement. Participants pressed SPACE when finished reading; then a Chinese prompt asked them to judge veracity by pressing 'F' (false) or 'J' (true) as quickly and accurately as possible. After all trials, overall accuracy rates for true and false statements were shown to participants. Data cleaning: Trials with reading time <300 ms or >20,000 ms were removed (3.9% of data). For probe responses, RTs <300 ms or >3,000 ms were removed (3.8% of data). Statistical analysis: Linear mixed-effects models (LMMs; lme4) were used for reading times and response times; generalized LMMs (GLMMs) for accuracy. Full models included fixed effects of content characteristics (veracity, length, place, time, source, calls to action); reduced models retained veracity, significant predictors, and interactions. Demographic models added age, family SES, and years of study as fixed effects. Subjects and items were random effects. Categorical predictors used sum coding (True = 1, False = −1). Models were fit by REML; model comparisons used χ²-distributed likelihood ratio tests, with AIC/BIC for fit. Post hoc contrasts employed emmeans with Tukey adjustment.

- Reading times: Only length significantly increased reading times (p < 0.001). No significant effects of veracity, place, time, source, or calls to action; no veracity × length interaction. - Accuracy: Overall accuracy was similar for true (M = 0.76, SD = 0.45) and false rumors (M = 0.76, SD = 0.46), t = −0.15, p = 0.879, d' = 0.003. Length reduced accuracy (p = 0.012) and interacted with veracity (p = 0.045): participants were less accurate on longer false rumors, but length did not reduce accuracy for true rumors. - Response times (RTs): Participants responded faster when correct than when wrong (LSMEANS: 727 ± 24 ms vs. 838 ± 20 ms; z = 9.85, p < 0.001). For correct judgments, they were faster rejecting false rumors than accepting true rumors (false: 696 ± 25 ms; true: 757 ± 26 ms; z = 3.80, p < 0.001). For wrong judgments, RTs did not differ between false and true rumors (false: 860 ± 28 ms; true: 816 ± 28 ms; z = −1.88, p = 0.237). Length interacted with judgment for RTs, but length itself did not slow correct RTs in the demographic-augmented model. - Demographics: Family SES significantly predicted higher accuracy (p = 0.009 in models including demographics). Age and years of university study did not significantly predict reading times, accuracy, or response times. Reading times showed no significant demographic effects; RTs showed no significant main demographic effects.

The findings address the research questions by showing that a linguistic characteristic—statement length—robustly increases reading time and reduces accuracy (particularly for false rumors), while most other examined content features (place/time/source/calls to action) did not affect processing or accuracy in this task. Demographically, higher family SES was associated with better truth discernment, aligning with prior survey research linking SES and misinformation susceptibility, and suggesting enduring SES-related differences in decision-making and information evaluation among university students. Processing-wise, participants were faster when correct and especially efficient at correctly rejecting false rumors, whereas accepting true rumors took longer. This pattern diverges from classic sentence processing findings that often observe slower responses for ‘false’ decisions, implying that rumor evaluation may involve different cognitive dynamics than standard truth-value judgments of well-established facts. The results highlight the importance of incorporating content features and socioeconomic background into theoretical accounts of rumor processing and inform practical interventions aimed at improving truth discernment.

Veracity judgment of COVID-19 rumors reflects an interaction between content features and reader characteristics. Longer statements take longer to read and, for false rumors, reduce accuracy, while true rumor accuracy is less affected by length. Participants are faster when making correct judgments, especially when rejecting false rumors. Higher family SES predicts better accuracy in veracity judgments among young adults. Future research should examine additional psychological variables, include more demographically diverse samples, and employ real-time measures (e.g., eye tracking) to capture moment-to-moment processing and to test how specific content manipulations influence both processing and truth discernment.

- Gender imbalance (predominantly female) precluded reliable assessment of gender effects. - Restricted sample (university students; narrow age distribution) limits generalizability and prevents robust age comparisons. - Task design required reading then post hoc judgment, limiting moment-to-moment processing inference; real-time methods (e.g., eye tracking) were not used. - Findings may not generalize to non-student populations or to other languages/platforms beyond the Chinese context and piyao.org.cn-sourced items.