Introduction

Inthe mobile internet era, people can engage in hot events and share their opinions, which also fostering an environment prone to online rumors. Especially, during the outbreak of COVID-19, the world has been fighting both a pandemic and a “massive infodemic“1. Researchers pointed out that this pandemic has changed people’s behaviors related to information sharing, including how rumors, misinformation, and fake news are shared2,3,4,5,6,7. Rumors mean information and news without confirmation or certainty about facts8,9. Sharing messages whose veracity is unknown to users may cause adverse impacts, such as the stockout of Shuanghuanglian in China10and panic purchasing of groceries and paper products in the United States11. Often, information with attractive headlines/hot topics or updates associated with breaking news stories can easily attract people’s attention and be shared them thousands of times, even possibly unsubstantiated or false content, some of which may be verified false8,9. For example, rumors collected by People’s Daily12. Rumors have proven very dangerous in the digital ecosystem and affect individuals, organizations, societies, and cultures13, which might have an essential social cost in the future.

The academic community has risen to challenge of investigating rumors sharing behavior2,3,14, rumor detection8,9, rumor rebuttal7,14, and how they spread to avoid their dramatic impacts. Some governments have recently launched investigations and implemented new rumors-sharing legislation15. Technology companies, such as WeChat and WhatsApp, have introduced security assistance in the platform to control the spread of false information and have also launched initiatives for anti-rumors. Previous research found that most social media users cannot identify whether the information is true or false13,16. Pennycook et al.16 found that 16.4% of people can accurately rate false information without tag control, and 53.0% can accurately rate truth information without tag control. The mean detection rate was 43.9% (median 45%), and only 17% of participants were better than chance (50%) in a research by Moravec et al.13. Rumors in social media have become a serious concern around the world. Rumors also attract the attention of academia from various disciplines15.

Understanding how humans evaluate credibility is a critical scientific question in the “massive infodemic” era. However, many studies have investigated rumor-related behaviors, attitudes, and methods to fight online rumors. These studies mainly investigate rumor detection based on subjective declarations2,17,18. Simply asking people whether they can detect rumors’ credibility is insufficient to reveal the real reasons for such a decision19. Psychophysiological tools (e.g., brain imaging tools) have recently received much attention due to their ability to complement other sources of data, such as self-reported data (i.e., questionnaires to investigate rumor-related behaviors, attitudes, and beliefs)16,20. Neurophysiological tools have proved valuable methods in understanding human responses in economics, psychology, and marketing20.

In contrast to other methods, the electroencephalogram (EEG) can record brainwaves at small intervals, even up to 20,000 times per second, and is cost-efficient compared to fMRI20,21. Dimoka20coined the term NeuroIS and advocated using neurophysiological tools in information system research with the advantages of not being susceptible to subjectivity bias, social desirability bias, and demand effects. EEG typically uses multiple electrodes placed on the scalp to detect potentials generated by neurons22, is cheaper than fMRI, is tolerant to movement, is silent, and is not constrained by the environment20. ERPs (event related potentials) are evoked by repeated stimuli and extracted from EEG23. ERPs are categorized based on their polarity (Positive or Negative) and timing, like N400 (a negative response occurring approximately 400 ms after the stimulus is presented). ERPs can measure real-time neural activities in the cognitive process and are widely used to explore human responses when people engage in various activities22,23.

Many previous studies used EEG (Electroencephalogram) or fMRI (functional magnetic resonance imaging) in the context of credibility focused on deception24or true and false recognition (conscious awareness in memory)25or continued influence of misinformation26,27. To the best of our knowledge, the present study also provides one of the first attempts to investigate the temporal dynamics of brain activity during online rumors recognition. So far, little research has been reported in the scientific literature that directly attempted to investigate brain activity involved in rumor detection, especially exploring the brain activity elicited by online rumors. In a recent article19, researchers used EEG to measure brain activity and applied Logistic Regression to predict human source credibility evaluations with an F1 score exceeding 0.7. In another work13, collected behavioral and EEG data to test whether people could detect fake news on social media and measure and confirm the effect of confirmation biases on fake news evaluation. Online rumors recognition is associated with trustworthiness, can be regarded as making decisions under uncertainty19. Online rumors recognition might be considered as a reasoned and deliberative assessment.

To track the time course of trust and making decisions under uncertainty, early and later ERPs components related to rumors processing were considered in this study. Specifically, we focused on frontal P2 that reflects emotion or attention capture28and was found to be influenced by negative stimuli29. At subsequent latencies, we focused on late posterior P3, peaking around 300 ms, that has been found to be related to information processing30, emotional and motivational processes31, and trust choices32. Finally, we considered the posterior negativity (N400) and the late positive potential (LPP). The N400 has long been thought of as a language measure and semantic memory, typically elicited in response to semantic priming and language processing, and found to be more negative in the thematically incongruent sentence33. Kutas and Federmeier33also pointed out that N400 can be evaluated as a nonspeeded, end-state response following a late positivity. LPP, peaking on centroparietal areas around 600 ms, signals an attentional response to motivational stimuli31,34, and the arousal and valence properties of emotional35,36. As rumors recognition is similar to trust assessment or credibility evaluations, we expected to observe higher or lower ERPs responses to outcomes following truth judgements than to outcomes following false judgements.

The diffusion of online rumors is fast and has been shown to affect our lives significantly. The extent to which people can distinguish between truth and false online rumors is still being determined. Moravec et al.13 pointed out that there is little empirical evidence about their ability. The second question is: How do users detect rumors when they receive rumors outside of their area of expertise and experience (typical of many rumors, like the example of anti-inflammatory drugs)? In this study, we examine people’s self-reported belief of rumors and use ERPs to describe and understand brain activity during rumor detection. Eighty rumors (half of them are true) were used in the experiment. Participants were required to detect the rumors displayed on the screen. At the same time, brain signals were collected to investigate the cognitive process during rumor detection.

Methodology

Participants

G*power 3.137was used to calculate the sample size. A minimal size of 22 was required to reach a large effect size (f = 0.4) in a 2-way within-subjects ANOVA with α (error probability) = 0.05 and β (error probability) = 0.9538. Hence, forty-five healthy and right-handed students (23 males and 22 females, aged 18–25 years, age: 19.6 ± 1.25 years) from different faculties were recruited via university bulletins. These subjects are right-handed and have no background in neuroscience and psychology. Moreover, they have normal or corrected-to-normal vision without history of nerve disorders or mental illness. The subjects were told not to drink coffee and to have enough rest the day before the experiment. Data from four subjects were discarded due to the high number of artifacts in EEG data.

Additionally, six subjects were excluded due to problems with the EEG montage and excessive movement artifacts. Finally, thirty-nine participants (20 males and 19 females, age: 19.5 ± 0.93 years) with at least twenty valid trials per condition entered the analyses. They all signed the consent forms before the experiment and were paid 50 ¥. This study was approved by the ethics committee the Institute of Neuroscience and Cognitive Psychology of Anhui Polytechnic University. All procedures were conducted in accordance with the Helsinki declaration.

Materials

Here, rumors are defined as an information and news without confirmation or certainty about facts8. Eighty rumors were selected from “科学辟谣” (piyao.kepuchina.cn) and “较真” (vp.fact.qq.com), which are two famous of the third-party fact-checkers. Half of the stimuli were from stories that were verified as fabricated and entirely untrue. Half of the stimuli were selected by choosing contemporary topics that did not contain factual errors or fabrication. There was no tagged warning on each of the stimulus. Participants needed to assess their credibility. The headlines were actually true or false and covered topics of COVID-19, health, technology, and food. The experiment mimicked the format of social media display. Stimuli were presented in a standard size of 800 × 600 pixels with a picture, headline, and details on the picture (see Fig. 1). Information about stimuli can be found in the supplemental material (https://pan.baidu.com/s/1wWV3hetrSWwcLLXoZXE53A, access code: 1234).

Fig. 1
Fig. 1
Full size image

A sample of information is shown to participants. They need to judge the truth of the presented information.

Procedure

The stimulus display screen was an Acer P229HQL screen with 1920 × 1080 resolution. E-prime 3.0 (Psychology Software Tools, Inc., Pittsburgh) was used to control the experimental program and materials. The subjects sat 65 cm in front of the screen. The visual angle is 13.5°× 7.5°. The sequence of stimulus presentation was randomized to ensure the fairness and reliability of the experiment. Every picture was repeated 30 times, and the representation time of the stimuli was 3000 ms. Then, experimenters introduce the experiment to the subject and put the cap on the subject’s head. The participants conducted an exercise before the experiment and were required to detect the truth of the rumors presented on the screen. The participants were asked to click the left mouse button if they judged the rumors true and click the right mouse button if they judged the rumors false. In addition, they click the middle mouse button if they cannot judge the truth of rumors. The participants did not know whether their responses were correct or incorrect. The procedure is shown in Fig. 2.

Fig. 2
Fig. 2
Full size image

The process of experiment.

Electrophysiological recording and ERP analysis

The 64 Ag/AgCl electrodes distributed according to the 10–20 international system were applied in the experiment (Fig. 3) and mounted in an elastic cap (Braincap, Brain Products GmbH, Germany). The brain signals were recorded using the Brain Vision Recorder software (Brain Products GmbH, Germany). Vertical electro-oculographic activity (Fp1) was recorded with additional electrodes located 1.5 cm above the left eye, and horizontal electro-oculographic activity (Fp2) was recorded with 1.5 cm outside the outer canthi of the right eye. All electrodes were referenced to the FCz site with a common ground (Gnd). Impedance levels of all the electrodes were set below 5 kΩ. Moreover, the brain signals were recorded at 1000 Hz and filtered with a bandpass of 0.05–70 Hz.

Fig. 3
Fig. 3
Full size image

The 64 Ag/AgCl electrodes used in the experiment.

Offline data was performed in Matlab R2018b (v.9.5; The MathWorks, Inc.) using EEGLab (version 2019.0), an open-source toolbox developed by Delorme and Makeig39. The EEG signals were re-referenced to the average of the bilateral mastoid and down-sampled to a sample rate of 500 Hz. The averages per channel were filtered with 0.01–30 Hz. In order to correct detailed artifacts, we used MARA (multiple artifact rejection algorithm), which is an open-source EEGLAB plug-in40,41, to clean the EEG signals. MARA has been proven effective in recognizing and rejecting eye- and muscle artifacts40,41. EEG signals were computed using EEG epochs that started from 200ms before stimuli onset and 1000ms after stimuli onset. Each epoch was baseline corrected using the signal during 200ms that preceded the onset of the stimulus. Moreover, those trials with an amplitude exceeding ± 75 µV were deleted. Event related potentials were then averaged separately for each channel and stimulus type (i.e., trails identified as true or false were averaged, and each condition had at least 20 trails). The event related potentials analyses were conducted on the mean amplitude values for specific sets of electrodes within predefined time windows. The single-subject ERPs were then used to derive the grand averaged waveforms for display and analysis. The mean amplitude was calculated for each component with a narrow time window centered around the components’ peaks in the grand average waveform42. According to earlier studies and best capture differences between conditions, mean ERP amplitude was measured for each response type for evaluation of the P2 (140–180 ms), N400 (320–420 ms), P3 (380–480 ms), and LPP(580–780 ms). The details of recording and analysis of brain signals are shown in Fig. 4.

Fig. 4
Fig. 4
Full size image

The process of Brain signal recording and analysis.

Statistical analysis

Data from headlines judged as true and false were compared using repeated measures ANOVAs (Response × Location) within-subject on the mean amplitude of each ERPs component. In addition, headlines related to COVID-19 were compared with other headlines. We also detect the gender difference. Paired t-test was used to determine the differences in subjective ratings and task performances between tasks. Violation of sphericity was handled with a Greenhouse-Geisser correction, as well as the effect size (partial eta-squared values η2) reported for all ANOVAs. The data analysis was carried out by using R Software (Version 3.5.3). All statistical significance reported for all tests are 2-tailed and were considered significant if they were less than 0.05.

Results

Behavioral data

Two-tailed t-tests were conducted to analyze the behavioral results. The response time and the detection rate are shown in Fig. 5. There was no significant difference between the two kinds of detection results for the response time (Rfalse=1672.03 ms, Rtrue=1605.36 ms, t(38) = 1.66, p = 0.114). There was a significant difference between the information types for the response time (RCOVID-19=1646.11 ms, Rothers=1605.26 ms, t(38) = 2.51, p = 0.020). There was no significant difference between the information types for the detection rate (RCOVID-19 = 0.41, Rothers=0.43, t(38)=−1.30, p = 0.215). There was no significant difference between males and females in response time and accuracy (Table 1). The mean detection rate was 42.2% (median 43.4%) and worse than chance (50%). Moreover, only 22% of our participants were better than chance.

Fig. 5
Fig. 5
Full size image

Behavioral results are shown as violin plots. (A) Violin plot quantitative analysis of response time from two response types. (B) Violin plot quantitative analysis of response time from COVID-19 and other headlines. (C) Violin plot quantitative analysis of detection rate from COVID-19 and other headlines.

Table 1 Comparisons of response time and detection rate between males and females.
Full size table

ERP analysis

Figures 6 and 7 display the grand average ERPs for true and false response, and Figs. 8 and 9 show the topographic distributions of primary ERP differences. Participants were required to make credibility evaluations. Five outcomes were obtained, which are true rumors with true judgment, true rumors with false judgment, false rumors with false judgment, false rumors with true judgment and uncertain. However, the number of trials in each response category was insufficient for reliable averaging. In such cases, it is necessary to pool the ERPs across response categories to obtain sufficient trials. Hence, the comparisons between different categories were not analyzed in this study. The two within-subject factors were conditions (true versus false responses) and locations (regions of ERPs generated).

Fig. 6
Fig. 6
Full size image

Grand average waveforms for four representative electrodes on the P2 and P3 components (time windows are marked in the rectangle).

Fig. 7
Fig. 7
Full size image

Grand average waveforms for four representative electrodes on the N400 and LPP components (time windows are marked in the rectangle).

Fig. 8
Fig. 8
Full size image

Topographic distributions of P2 and P3 differences between true and false.

Fig. 9
Fig. 9
Full size image

Topographic distributions of N400 and LPP differences between true and false.

recognition for the specific time windows of interest (Coloration indicates voltage amplitude).

P2

In the first interval, a condition (judged-true, judged-false) × region (AFz, Fz, FCz, Cz) ANOVA was performed to confirm that the true/false recognition reported in previous ERP studies13,19,26,27,34. Staring at around 140 mm and extending until 180 mm, the waveforms elicited by truth identification were less positive than false. Statistical analysis revealed a main effect for the condition with F(1, 37) = 6.99, p = 0.012, η2 = 0.159, but there was no main effect for the region with F(2.0, 73.9) = 1.17, p = 0.325, η2 = 0.031. There was also a significant interaction effect between condition and region, F(1.8, 65.5) = 6.23, p = 0.001, η2 = 0.144. Post hoc t-tests revealed significantly more positive amplitudes for judged-false than for judged-true condition at electrode positions FCz (MD (Mean Difference) = 0.566 µV, SE = 0.162, p = 0.001), and Cz (MD = 0.825µV, SE = 0.180, p < 0.001) (Figs. 6 and 8).

An independent sample t-test was conducted to analyze the difference between males and females. A gender effect in the P2 was discovered. Larger P2 was evoked for males than females in the conditions (ps < 0.05) (supplementary data).

P3

Like P2, the repeated-measures ANOVA shows a main effect for the condition with F(1, 37) = 15.64, p < 0.001, η2 = 0.297, but there was no condition × region interaction (F(1.7, 63.3) = 2.63, p = 0.088, η2 = 0.066). For the P3, regardless of regions, the average amplitude significantly differed between the experimental conditions, indicating significant differences at all topographic regions (ps < 0.05) with larger amplitudes observed under “judged- false” conditions (Figs. 6 and 8).

An independent sample t-test was conducted to analyze the difference between males and females. The results showed that there was a significant difference between males and females. Larger P3 was evoked for females than males in all the regions (ps < 0.05) (supplementary data).

N400

For N400, a condition (judged-true, judged-false) × region (CPz, Pz, POz, Oz) ANOVA was conducted to confirm the true/false detection. The analysis revealed a main effect of the condition (F(1, 37) = 21.67, p < 0.001, η2 = 0.369) and a condition × region interaction (F(1.8, 67.9) = 4.31, p = 0.020, η2 = 0.104). Post hoc t-tests revealed significantly more negative amplitudes for judged-true than for judged-false condition at electrode positions CPz (MD = −1.396 µV, SE = 0.231, p < 0.001), Pz (MD = −0.977 µV, SE = 0.296, p = 0.002), POz (MD = −0.882 µV, SE = 0.231, p = 0.001), and Oz (MD = −0.759 µV, SE = 0.241, p = 0.003) (Figs. 7 and 9). The N400 is believed to be associated with increased general cognitive ability on more effortful tasks of executive function, information processing, and verbal memory and learning.

An independent sample t-test was conducted to analyze the difference between males and females. The results showed a significant difference between males and females in the mean amplitude of N400. Enhanced N400 was evoked for males than females in all the regions (ps < 0.05) (supplementary data).

Late positive potential

Like N400, a 2 × 3 repeated measures ANOVA (condition × electrode) was used to analyze amplitudes of LPP. There was no main effect for the region (F(1.6, 59.5) = 2.66, p = 0.090, η2 = 0.067). There was a significant main effect for the condition (F(1, 37) = 13.33, p = 0. 001, η2 = 0.265) and a condition × region interaction (F(1.8, 69.0) = 9.16, p < 0.001, η2 = 0.198). Post hoc t-tests revealed significantly more positive amplitudes for judged-false than for judged-true condition at electrode positions CPz (MD = −1.293 µV, SE = 0.208, p < 0.001), Pz (MD = −1.286 µV, SE = 0.243, p < 0.001), POz (MD = −0.728 µV, SE = 0.320, p = 0.029), and Oz (MD = −0.333 µV, SE = 0.239, p = 0.173) (Figs. 7 and 9).

An independent sample t-test was conducted to analyze the difference between males and females. The results shown a significant difference between males and females in the mean amplitude of LPP. Larger LPP was evoked for females than males in all the regions (ps < 0.05) (supplementary data).

Discussion

The effect of rumors authenticity on behavioral response

Understanding how humans detect false information is a crucial scientific question in the era of rumors. The present study examined whether social media users were adept at distinguishing between true and false rumors and investigated brain activity correlates of true and false rumors detection using a similar paradigm introduced by Moravec et al.13 and Kawiak et al.19. Our results show that people are poor at separating false rumors from real ones, whether the rumors are related to COVID-19. People responded faster (not significant in statistics) when they judged the rumors as true than rumors judged as false. In addition, there was no gender difference in response time and accuracy. Our findings are consistent with a previous study by Moravec et al.13. The accuracy of distinguishing true rumors from false rumors is 42.2%, which is less than chance (i.e., 50%), yet only 22% of our participants were better than chance.

The effect of rumors authenticity on ERPs components

The ERPs results show two interesting findings. First, when participants viewed headlines that supported their beliefs, more negative N400 was evoked at central and parietal locations than headlines that opposed their beliefs. Moreover, ERPs for distrust in headlines were more positive at frontal, central, and parietal locations than headlines that supported their beliefs. Second, significant gender differences were found for participants in several ERP components (P2, N400, P3, and LPP) in the context of online rumors processes and recognition behavior. Males had enhanced P2 and N400 compared to females, while females displayed larger P3 and LPP amplitudes than males. Significant effects for rumors detection in both negative and positive ERPs components partially support the results of Kawiak et al.19. They found significant differences in ERP signals in two broad time intervals: 268 ~ 300 ms and 372 ~ 796 ms from trust and distrust decisions. Enhanced negative ERPs (N400) were evoked for trust decisions than distrust decisions, and larger positive ERPs (P2, P3, and LPP) were elicited for distrust decisions than trust decisions. However, the current study discovered a significant rumors detection and gender effect for all the ERPs components. In comparison, the previous study did not take gender into consideration and reported no ERPs difference between stimuli. The differences in the nature of the stimuli used may contribute to the bias in results. In their study, “Kanji signs” were presented with the question of whether the translation of that sign was correct. In contrast, rumors in this study were presented on the screen, which required the participants to determine whether the information was real or fake.

Our results demonstrate significant P2 effects, lending support to the idea that rumors’ authenticity can influence early attentional processes. The typical P2 component is located around the centro-frontal and parieto-occipital regions43. Increased P2 activity has been observed when focusing on threatening or negative stimuli29. In this view, the P2 amplitude probably represents the negative attitude elicited by false information, which indicates that P2 was sensitive to false information. Additionally, it has been suggested that the P2 may represent elements of higher-order perceptual processing influenced by attention, which may be related to cognitive matching44,45. Freunberger et al.44 found higher P2 amplitudes for incongruent (between prior information and future response) than for congruent. From this view, rumors incongruent with participants’ beliefs might cause negative attitude and evoke higher P2 in the early stage.

Also, the N400 effect, peaking around 400 ms at parietal regions, was discovered, where more enhanced N400 activity was elicited when participants judged the rumor as true or believed in the rumor than judged the rumor as false (or distrusted the rumor). The N400 has long been thought of as a language measure and semantic memory, typically elicited in response to semantic priming and language processing46. Here, an enhanced centro-parietal N400 was obtained when participants discerned true from false rumors, different from semantic incongruency. Our results suggest that N400 is not dependent on an information state such as “recognition” and is thus open to stimuli in many ways similar to the decision-making process. True information was “motivated by the desire to exclude confounding variables such as participant prior knowledge, experience or opinions on the subject of the evaluated message.”19. In this regard, a larger N400 was obtained for rumors that supported their beliefs. The data suggest increased activation in the trust condition (Fig. 9), perhaps indicating movement towards the true rumors13 but not cognition bias or incongruence between presented information and their cognition.

The results of this study showed that the distrust condition evoked a larger amplitude of P3 (380–480 ms) than the trust condition in frontocentral regions. P3 is the most-studied ERP component, which reflects selective attention and information processing30,47. Anterior P3 is sensitive to the amount of attentional resources allocated to a stimulus, increasing with greater focal attention47. For this view, participants paid more attention to rumors challenging their beliefs, which is inconsistent with13. They found that people paid more attention to headlines supporting individuals’ beliefs when seeing a fake news flag on a headline13. Here, rumors without flags were taken as stimuli, which may contribute to this bias in conclusion. Donchin and Coles48 pointed out that the P300 component may reflect people’s cognitive conflicts after recognizing prompt information and attentional resources allocated to the stimulus. That is to say, once individuals recognize that information challenges their a priori belief, they do not stop thinking about it. In contrast, an opposite result was obtained by Moravec et al.13. This inconsistency in the literature may be task-related. Rumors related to COVID-19 concern the public, while news related to US politics was used by Moravec et al.13. University students were recruited as participants in both studies. Our experiment was conducted during the pandemic, and participants are more sensitive to rumors related to COVID-19. Students who possibly do not care about topics related to US politics were recruited by Moravec et al.13. Moreover, this study required participants to respond in a shorter time.

Last but not least, we found that the distrust condition evoked a larger LPP (580–780 ms) than the trust condition in posterior regions. The LPP has demonstrated that can be observed throughout the scalp with a maxima over centroparietal areas34, and LPP amplitude is predominantly thought to reflect overt and post-perceptive attentional response to motivational stimuli31,34, and the arousal and valence properties of emotional35,36. The results indicated that after an involuntary allocation of attention to detect and categorize rumors in the 380–480 ms period, truth and false (trust and distrust) were considered for making a final decision. Figure 9shows that LPP is associated with activity in the left medial temporal and parietal regions. The LPP may be localized to the left parahippocampal gyrus and is associated with episodic memory retrieval and true and false item memory differentiation46. But true memories are associated with a higher P3 and LPP than false ones in Kiat and Belli’s study46. This inconsistency in the literature maybe is paradigm-related. Meek et al.24presented similar results to ours in a study of the misinformation effect. Larger P3 and LPP activity were obtained in evaluating misinformation-based items than in evaluating true information during the response selection stage24. Both Meek et al.24 and our studies investigated the neural activity evoked by true and false information (i.e., sentences but not a single word were used as stimuli), which may be the main reason for inconsistent research results.

Gender differences in rumors processing strategies

Gender-related differences were found in the amplitude of the four investigated ERP sub-components during rumors detection. Males exhibited more positive P2 and enhanced N400 compared to females. In comparison, females displayed larger P3 and LPP amplitudes in detecting the rumors than males. No gender differences were observed in behavioral responses. The results indicated that differences in behavioral response did not accompany the neural gender differences during rumor detection. It is most likely that ERPs were able to detect even minor changes at neutral levels that were undetectable at behavioral levels20,49.

The P2 is related to perceptual analysis and attention allocation, which reflects an early automatic stage44,45. Males had significantly higher P2 amplitudes than females in this study. Previous studies observed similar findings50,51. This indicates that males might devote more involuntary attention to the subsequent stimuli than females, reflecting enhanced reactive attention in the early stage. More negative N400 was evoked in parietal regions for males than females. Our results partially support the results of Jaušovec and Jaušovec52. More negative N400 might imply that men required more effort to process visual information, requiring extensive elaboration to reach a decision.

Females displayed larger P3 and LPP amplitudes in rumors detection than males, indicating that women tend to devote more sustained attention to rumor detection than men. Some previous studies have suggested that the anterior P3 reflects the cognitive evaluation of stimuli and is an attention-driven stimulus signal30,47. In addition, the LPP signals an attentional response to motivational stimuli43,53. In this regard, ERPs findings suggest a discernible influence of sex on attention. In the context of rumor detection, men seem to develop an enhanced top-down process of attention bias toward the truth or false rumors. High-order, top-down regulatory processes may modulate fundamental perception, attention, and response54. Thus, men allocated more neural resources after stimulus presentation, perhaps using a mechanism to respond faster, which mainly exhibits larger P2 and N400. While more proactive and cautious cognitive processing in women was applied, which reflects larger P3 and LPP (both indexes sustained attention). Lim and Kwon’s survey55also supports our findings, and they pointed out that women exercise more caution concerning the quality of information on online platforms. As a result of these findings, it is suggested that top-down and bottom-up processes interact to generate different strategies for information processing between the sexes based on task demands54,56. It is notable in our work that sex differences in temporal dynamics emerged in distinct stages of information processing related to brain activity, possibly due to different strategies employed.

In addition, other psychological and sociological perspectives might contribute to the gender difference, though we did not collect these data. The sample consists of university students, which made the background characteristics of participants (same education level, age, major) were controlled. While other characteristics of participants may contribute to the gender-related difference. Ibrahim et al.57found that people with high agreeableness tend to believe rumors than people with low agreeableness, and there was a significant correlation between believing rumors and personality. In this view, personality difference may contribute to the gender difference in ERPs components. In addition, their ability to identify false content might also affect the gender-related difference58.

Implications and limitations

Theoretic and practical implications

First and foremost, future research needs to investigate why people cannot distinguish and believe in rumors. Our results support that “most social media users would make better truth judgments by flipping a coin.”13. The experiment was conducted during the COVID-19. Moreover, participants had a utilitarian mindset during the experiment. However, Johnson and Kaye59 pointed out that social media users are often in a hedonistic mindset, and individuals may think more deliberately about the information than in daily usage scenarios. Thus, the real problem may be worse. Hence, future research is needed on how to fight rumors and why they believe even when they make no sense at all. Second, to the best of our knowledge, the present study also provides one of the first attempts to investigate the temporal dynamics of brain activity during rumor processing during the COVID-19 pandemic. It would be useful to repeat the same experiment using brain imaging equipment to explore in further detail how the human mind processes rumors and information against their beliefs.

Our results indicate that individuals are poor at distinguishing between true and false rumors. One possible application would be the credibility evaluation of debunking false rumors designed to fight rumors. Our study finds the temporal dynamics of brain activity during rumor detection. In particular, the P2, P3, N400, and LPP can indicate whether people believe in rumors. In addition, men and women showed different rumor-processing mechanisms in our study. Hence, in the near future, we can develop a prompt system to prevent being deceived by rumors based on their brain responses.

Limitations and future studies

Like other studies using neurophysiological data, this study has the same limitations. We conducted the study in a lab experiment to carefully control exogenous factors. Participants were wearing electrode caps before a screen, and rumors were displayed on separate pages instead of scrolling homepages using smartphones in a hedonic context. We introduced the procedure before the experiment, which may trigger a utilitarian mindset of thinking more deliberately about the information displayed than in the usual setting of social media use13.

Nevertheless, the utility of the ERP data for understanding the rumor recognition behavior in this study is limited by the fact that there is no evidence of any association between ERPs and behavior. Thus, more research will be needed to understand the neural locus of rumor recognition as manifest in behavioral responses. In addition, we found that only 22% of our participants could detect false rumors better than chance. As Moravec et al.13said, “Most social media users would make better truth judgments by flipping a coin.” Then, what should we do? Does a false rumor flag affect judgments about truth? If not, what should we do to correct people’s cognitive biases? Hence, there is a need for more research on how to improve users’ ability to discern truth from fiction in social media and more research on how to entice them to invest more time and attention in the information they see and to restrain online rumor sharing behavior without checking the truth13.

Additional limitations come from the homogeneous sample of undergraduate students with neuroplasticity prevalent in adulthood, and the features of ERPs components change with brain maturation60. Therefore, we must be aware of studies investigating the underlying neurophysiological difference between young and elderly groups.

Conclusion

The escalating infodemic of COVID-19-related rumors presents an immense societal challenge. This study aimed to shed light on individuals’ ability to discern between true and false rumors in the context of COVID-19, investigating the associated neural processes. Our findings underscore the susceptibility of individuals to false rumors, with only 22% of participants demonstrating the capability to distinguish rumors better than chance. Furthermore, our results highlight the utility of EEG-derived components, including P2, P3, N400, and LPP, as potential indicators of trust or skepticism during rumor detection. Notably, when individuals exhibited distrust in rumors (judging them as false), we observed the elicitation of larger positive ERP components (P2, P3, and LPP), whereas trust in rumors led to enhanced negative ERP responses (N400). These ERP findings also illuminate the presence of sex-related disparities in the temporal dynamics of information processing, likely stemming from divergent cognitive strategies. Consequently, this research underscores the potential of EEG as a valuable tool in extending our understanding of rumor processing, complementing traditional data sources such as surveys and interviews. In closing, our study underscores the imperative need to combat pandemic-related rumors on social media platforms in the future.