Emotional drivers of misinformation consumption and toxicity on YouTube

Introduction

Social media platforms are central venues for sharing information and emotions in digital form (Schmidt et al., 2017; Del Vicario et al., 2016; Stella, Ferrara & De Domenico, 2018). Every interaction–whether a post, like, share, or reply–leaves behind a digital trace that encapsulates users’ communicative intentions and cognitive-emotional states (Stella, 2022). The scale of these traces provides researchers with unprecedented datasets to examine online discourse and social dynamics through computational approaches (Cinelli et al., 2020; Santoro et al., 2023). In these highly emotional environments, information processing can vary widely. Users may engage with content analytically or react emotionally (Pennycook & Rand, 2021; Pennycook et al., 2020; Pytlik, Soll & Mehl, 2020). Constant exposure to large volumes of often low-quality content can lead to “social media fatigue” (Ahmed & Rasul, 2023), a well-documented phenomenon of distress caused by the need to internalize and decide upon (e.g., reshare/like) vast amounts of information.

Users cope with this distress in various ways (Ahmed & Rasul, 2023), including the emergence of toxic communication–hostile or aggressive language that violates norms of civility (Cinelli et al., 2021; Zollo et al., 2015; Avalle et al., 2024). Such toxicity can result from users’ inability to regulate their emotions while communicating with others (e.g., emotional venting (Vermeulen, Vandebosch & Heirman, 2018)) or from a lack of cognitive resources to engage in analytical thinking when confronted with incorrect or misleading content, i.e., misinformation (Pennycook & Rand, 2021, 2019).

While misinformation is often framed as a consequence of cognitive failure (McIlhiney et al., 2023; Pantazi, Hale & Klein, 2021), recent research challenges the simplistic notion that misinformation-prone (MIp) users are merely less intelligent or “lazier” than those aligned with mainstream-affiliated sources (Ahmed & Rasul, 2023; Ecker et al., 2022; Borukhson, Lorenz-Spreen & Ragni, 2022; Taurino et al., 2023).

In this study, we adopt a data-driven approach to examine the emotional and toxic content of user comments on Italian YouTube videos related to COVID-19, a topic that has provoked intense polarization and emotional engagement (Cinelli et al., 2020; Santoro et al., 2023; Chen, Lerman & Ferrara, 2020; Polizzi, Lynn & Perry, 2020). Existing studies often overlook the interaction between misinformation, toxicity, and emotional language use. Our analysis focuses on two distinct user communities: misinformation-prone users (MIp), who predominantly engage with sources known for spreading false or misleading content, and mainstream-affiliated users (MAp), who primarily interact with reliable sources. This distinction enables a comparative analysis of how these groups express emotion and toxicity, and whether different informational environments are associated with distinct emotional patterns.

To investigate emotions we employ the recent computational framework of EmoAtlas (Semeraro et al., 2025), powered by Plutchik’s psychoevolutional theory of emotions (Plutchik, 1982, 2001). This approach identifies eight basic emotions—fear, anger, sadness, disgust, anticipation, surprise, trust, and joy—and was selected over other frameworks such as the circumplex model of affect (Posner, Russell & Peterson, 2005) due to: (i) its availability through large-scale, psychologically validated datasets (Mohammad & Turney, 2013), and (ii) its recent validation against human-annotated data via interpretable statistical analyses (Semeraro et al., 2025).

Importantly, in Plutchik’s theory, basic emotions can co-occur to form more nuanced “secondary” emotions or dyads (Plutchik, 1982; Mohammad & Turney, 2013). For instance, joy conveys positive reinforcement and pleasure, while trust signals safety, security and openness. When joy and trust co-occur, the resulting combined experience can feature, at the same time, deep connection and openness, which are traits characterizing love. In this sense, according to Plutchik’s psychoevolutionary theory, love can be considered as a dyad between trust and joy. We refer to the Table S2 for the full list of secondary emotions, as organized in Plutchik’s theory.

Our objectives are twofold: (i) to develop a taxonomy of emotions expressed by users oriented toward misinformation and mainstream news, across varying levels of language toxicity (as defined in Cinelli et al. (2021)); and (ii) to identify the emotional dyads or features most prevalent within mainstream-prone and misinformation-prone user categories. To this end, we introduce a novel analytic framework that integrates insights from cognitive psychology, network science, and explainable artificial intelligence (AI). At the core of this framework is the concept of arborescence, a measure of how secondary emotions or dyads derive from shared root emotions in user comments. Lower arborescence suggests emotional convergence around a dominant basic emotion, while higher arborescence indicates greater emotional diversification.

To guide our investigation, we articulate four research questions that address key dimensions of emotional and toxic language use in user discourse. First, we examine the emotional characteristics of toxic comments in response to content from Italian YouTube sources differing in credibility (reliable vs. questionable)[RQ1]. Second, we investigate how misinformation-prone and mainstream-prone users differ in their expression of Plutchik’s basic emotions [RQ2]. Third, we explore differences in the use of Plutchik’s emotional dyads across these two user communities [RQ3]. Finally, we assess whether the emotional framing of discourse varies systematically between misinformation-prone and mainstream-prone users, potentially reflecting distinct affective and cognitive environments [RQ4].

Our findings reveal that mainstream-prone users tend to express a broader range of emotional dyads, indicating higher emotional arborescence and greater affective nuance. By contrast, misinformation-prone users’ discourse is often centered around a narrower emotional set, particularly the combination of trust and fear–suggesting a more constrained and focused emotional landscape.

Related work

Following the framework proposed by George, Gerhart & Torres (2023), research on emotions and misinformation can be categorized into three overarching areas: (1) Stimuli, which focus on the triggers and motivations for engagement; (2) Actions, which encompass the processes of fabrication, propagation, and response; and (3) Outcomes, which investigate the effects and consequences of misinformation. We use this categorization to better contextualize the works discussed below.

Prior studies situated mainly within the “Actions” category have shown that misinformation tends to be characterized by narrower language use and reduced emotional richness compared to factual content (Carrasco-Farré, 2022; Ghanem, Rosso & Rangel, 2020). Other works spanning both “Actions” and “Outcomes” have highlighted the central role of emotional language in the spread and detection of misinformation. Exposure to false content has been found to elicit elevated levels of anger, fear, and resentment (Farhoudinia, Ozturkcan & Kasap, 2024; Cosgrove & Bahr, 2024), while anger in particular increases the likelihood of sharing misinformation (Han, Cha & Lee, 2020). In contrast, factual news is generally associated with more diverse and positive emotional responses (Farhoudinia, Ozturkcan & Kasap, 2024; Carrasco-Farré, 2022). A growing body of work has also explored how conspiratorial narratives rely on narrow emotional framing and emphasize perceived threats (Cosgrove & Bahr, 2024), while reduced emotional and lexical diversity has been linked to low-complexity communication in misinformation (Ghanem, Rosso & Rangel, 2020; Carrasco-Farré, 2022). Although these studies have advanced our understanding of misinformation’s affective traits, they predominantly focus on the content itself rather than user responses, and are thus primarily aligned with the “Actions” domain.

Our study shifts this perspective by focusing on user reactions to misinformation and mainstream content on YouTube, positioning our work closer to the “Stimuli” and “Outcomes” categories. Specifically, we examine emotional expression in user comments, not just in terms of individual emotions but also through a novel metric—emotional arborescence—that captures the richness and structure of emotional co-occurrence. This moves beyond the label-level analysis typical in previous studies and introduces a new way of quantifying emotional complexity in user discourse.

In addition to emotional content, our analysis incorporates toxicity, drawing on prior work that links it to online polarization and intergroup hostility (Cinelli et al., 2021; Mosleh, Cole & Rand, 2024). However, the relationship between toxicity and emotional expression remains underexplored. By jointly analyzing emotion and toxicity, we provide a more nuanced view of how users communicate in response to reliable vs. questionable sources.

A further contribution of our study is its user-centered design. Most existing work relies on aggregated analyses at the level of individual posts or videos. Our dataset includes multiple comments per user, allowing us to examine affective patterns not only in aggregate but also at the individual level. This enables us to observe emotional dynamics both within users and across the broader information ecosystem. By combining emotional arborescence, toxicity analysis, and user-level observations, our approach enables the investigation of how users engaged with different types of content (i.e., misinformation vs. mainstream) express emotions in measurably distinct ways—not just in which emotions dominate, but also in how they interact and diversify. In this way, we extend prior findings and introduce new tools for understanding emotional expression across polarized online environments.

Methods

In this section, we present the data and methodologies used to analyze the emotional content of texts in our data sources. We begin by detailing the variables and data processing steps. Next, we illustrate key concepts and definitions related to misinformation, toxicity, and the Plutchik emotional framework. Finally, we describe the techniques employed to assess the emotional content of the documents.

Results

We begin this section by presenting the distribution of emotions in comments, examining their relationship to both the trustworthiness of the sources and the toxicity of the language used. We then explore the emotional profiles of content from mainstream-prone and misinformation-prone users within the Plutchik framework, considering both basic emotions and emotion dyads. Finally, we investigate the overall emotional variety in the discourse of the two user groups.

Comments’ emotions in relation to toxicity and misinformation

We start our analysis by measuring the frequency of occurrence of Plutchik eight basic emotions (see Methods for details).

Thus, we label each comment to reliable and questionable sources with eight binary indicators to signal the presence or absence of each basic emotion.

We observe that $705,856$ comments ( $55.4\mathrm{\%}$) contain at least one basic emotion. To simplify our analysis, we exclude comments that do not exhibit any basic emotion. To ensure that this exclusion does not affect the proportion of comments from reliable and questionable sources, we conduct a ${\chi }^2$ test. The results indicate no significant relationship between the source’s trustworthiness and the presence of emotional features (Pearson ${\chi }^2(df=1,N=1,273,929)=16.462,p\text{-}\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{u}\mathrm{e}=0.975$). This confirms that removing comments without emotional features does not impact the distribution of trustworthiness labels.

However, the test does not account for the nature of emotional features. To explore this further, we compare the frequency of each distinct basic emotion in comments from reliable and questionable sources, as illustrated in Fig. 1. We then apply the ${\chi }^2$ test to each basic emotion individually. As shown in Table 4, significant differences emerge in the distribution of labels for all basic emotions except sadness, although the effect sizes, measured by Cramér’s V, remain low. A possible interpretation for the low effect size is that it reflects both the imbalance between the two groups and the pairwise nature of Cramér’s V: single emotions have limited impact, while combinations of emotions—which could carry stronger signals—are not captured. This indicates that following analyses should consider the interaction between emotions.

Table 4:

Pearson’s ${\chi }^2$ test and Cramér’s V on emotions in comments from videos from YouTube channels with different reliability labels.

Pearson’s ${\chi }^2$ test have $1df$, and Cramér’s V impact size on significant emotion presence in comments with differing channel reliability labels are computed with $0.975$ confidence interval. Significant ${\chi }^2$ statistics (p-value $<0.001$) are shown in bold. All emotions except Sadness show significant ${\chi }^2$ values but small impact size.

Emotion	N	$ch{i}^2$	Cramér’s V
Trust	293,961	31.96	0.007 $\pm$ 0.002
Joy	218,203	1,125.15	0.04 $\pm$ 0.002
Anticipation	168,156	55.45	0.009 $\pm$ 0.002
Surprise	96,246	242.23	0.019 $\pm$ 0.003
Fear	173,180	642.4	0.03 $\pm$ 0.003
Sadness	195,885	8.86	0.004 $\pm$ 0.002
Anger	122,948	236.87	0.018 $\pm$ 0.002
Disgust	97,189	46.92	0.008 $\pm$ 0.002

DOI: 10.7717/peerj-cs.3398/table-4

Next, we examine the relationship between emotions, toxicity, and source trustworthiness. Each comment in our dataset falls into one of four toxicity categories, as defined in subsection Toxicity Labeling of the Methods section: Acceptable, Inappropriate, Offensive, and Violent.

To answer RQ1, and investigate how toxicity levels relate to emotional features, we compute the probability of a significant presence of each basic emotion $e$ in a comment with toxicity level $t$ from a source with trustworthiness $r$, expressed as $P(v_e|t,r,{\bigcup }_{m\in E}{v}_m)$. Here, $v_e$ represents the presence of a significant signal for emotion $e$, while ${\bigcup }_{m\in E}{v}_m$ indicates that we consider only comments exhibiting a significant signal for at least one emotions $m\in E$.

As illustrated in Fig. 1, a clear relationship emerges between emotions and toxicity levels. Emotions can be broadly grouped into two categories: “positive” emotions (trust, joy, anticipation, and surprise), which negatively correlate with an increase in toxicity, and “negative” emotions (fear, anger, sadness, and disgust), which positively correlate with an increase in toxicity.

Furthermore, we find that “positive” emotions are more prevalent in low-toxicity comments from reliable sources compared to questionable sources. In contrast, “negative” emotions occur more frequently in high-toxicity comments from questionable sources. This finding is not unexpected, since toxicity and negative emotions are tied. However, here we find that anger is more likely to appear using appropriate language in comments to unreliable sources rather than in comments to reliable ones. This trend aligns with the expected pattern of other negative emotions. These findings address RQ1 and underscore how emotional signals vary in relation to both the toxicity of comments and the trustworthiness of their sources.

User emotional features and misinformation

In the previous section, we observed differences in the emotional expression of comments to reliable and questionable sources. Here, we focus instead on users and their preference for commenting on questionable or reliable sources (regardless of their personal beliefs. We employ the same categorization of users employed in Cinelli et al. (2021) based on their commenting preferences: those predominantly engaging with reliable YouTube channels (mainstream-prone), and those primarily engaging with questionable YouTube channels (misinformation-prone) and excluding users with undefined commenting preference (see Methods section Users’ Defining Features for details). Additionally, we have to exclude users without any significant emotional signal in their comments, as emotional features are missing. However, misinformation-prone users using emotions in their comments are 2.3% more numerous than mainstream-prone users, which is significant according to Pearson’s ${\chi }^2$ test with Yates continuity correction ( ${\chi }^2(df=1,N=313,680)=129.1,p\text{-}\mathrm{v}\mathrm{a}\mathrm{l}\mathrm{u}\mathrm{e}=2.2\times {10}^{-16}$, Cramér’s V: 0.018). Lastly, we focus on users posting a minimum threshold of $T=8$ comments, under the premise that these users had ample opportunities to express their emotions. Nonetheless, we performed robustness tests using thresholds of $T=5$ and $T=10$, finding similar results.

After training a random forest model which uses users’ emotional features to label their commenting preferences, we compute the model’s Shapley values, summarized in Fig. 2 (further details are provided in Methods section, Users’ Defining Features, the mean absolute SHAP values are reported in Table 5). Answering RQ2, we find that trust and fear are positively correlated with a user’s preference for questionable content: trust is a positive emotion associated with affiliation and sense of belonging; fear on the other hand is a negative emotion associated with the reaction to flee in response to a threat, in opposition to the drive to attack implied by anger. This finding is counterintuitive, as trust is a more prevalent in reliable comments, and suggests that trust, alongside fear, plays a key role in characterizing the language of misinformation-prone users. This suggests the possibility of an interplay between trust and fear, which gives rise to the dyad of “submission” in Plutchik’s theory (Plutchik, 1982). An alternative interpretation is that trust is indeed characterizing misinformation-prone users, as indicated in past works (Ognyanova et al., 2020), while other emotions are more prevalent mainstream-prone users.

Table 5:

Mean Absolute SHAP values for the random forest model.

Emotional feature	Mean absolute feature value
Trust	0.13641
Joy	0.13051
Fear	0.12127
Sadness	0.11967
Anger	0.11790
Anticipation	0.11442
Surprise	0.10433
Disgust	0.10093

DOI: 10.7717/peerj-cs.3398/table-5

Emotion co-occurrences

The difference highlighted by the SHAP value analysis in the use of emotional language between the misinformation-prone and mainstream-prone users classified by our model considers the impact of the eight distinct basic emotions. However, this SHAP-based approach neglects the simultaneous presence of pairs of emotions, i.e., dyads. To address this issue, we compute the probability $P(v_e,v_m|C_u)$ of finding the dyad $(e,m)$ in comments from user $u$ for each of the $28$ dyads. Figure 3 shows the differences in average values for the two user groups.

A two-part testing strategy is required to assess the significance of the values reported in Fig. 3, given the abundance of zeroes in the distribution of dyad values. We first test if the presence of dyads is independent of the user commenting preference for reliable or questionable sources; then, we test for the difference in intensity of the expressed dyads, i.e., how many comments from each user contained each dyad.

The results, fully reported in Tables 6 and 7, show a complex picture: in general, mainstream-prone users tend to employ dyads more frequently than misinformation-prone users, as all dyads with significant p-values are more frequent in mainstream-prone users. However, misinformation-prone users who employ dyads do express them more frequently.

Table 6:

Identifying significant differences of emotional co-occurrences between user groups.

Mann–Whitney U test results on non-zero user dyad values with FDR p-value adjustment. Rows where p-value $<0.05$ are in bold. The test confirms the significance of the dyads at the top of Fig. 3 of the main article, i.e., the dyads whose difference in the average values are higher for misinformation-prone users.

Dyad	MAp median	MIp median	W statistic	p-value
frozenness	0.077	0.100	1,200,735.5	<0.001
submission	0.059	0.077	618,883.5	<0.001
dominance	0.045	0.067	244,400.5	0.001
shame	0.067	0.077	903,541.0	0.028
ambivalence	0.042	0.067	140,593.0	0.001
aggressiveness	0.044	0.053	245,429.5	0.111
contempt	0.071	0.083	1,039,114.5	0.111
alarm	0.045	0.056	259,267.0	0.028
unbelief	0.038	0.047	90,961.5	0.111
sentimentality	0.062	0.077	643,625.0	0.019
outrage	0.043	0.053	181,567.5	0.028
pride	0.038	0.056	81,176.5	0.086
hope	0.091	0.100	1,813,266.5	0.071
envy	0.078	0.091	1,189,395.5	0.028
guilt	0.043	0.061	175,291.0	0.028
confusion	0.069	0.077	901,055.5	0.026
anxiety	0.059	0.067	538,413.0	0.049
morbidness	0.040	0.044	104,142.0	0.617
cynism	0.038	0.044	96,059.0	0.276
despair	0.100	0.111	1,832,459.0	0.060
disappointment	0.050	0.059	331,368.0	0.086
bittersweetness	0.054	0.059	355,884.0	0.213
remorse	0.077	0.083	1,140,266.0	0.111
love	0.100	0.111	2,228,767.5	0.291
pessimism	0.059	0.071	536,727.0	0.221
curiosity	0.067	0.075	627,373.0	0.099
optimism	0.083	0.086	1,577,441.0	0.344
delight	0.071	0.071	928,614.5	0.577

DOI: 10.7717/peerj-cs.3398/table-6

Table 7:

Comparison of emotional dyad prevalence between user groups using the Chi-squared test.

${\chi }^2$ test results ( $df=1,N=25,153$) with Yates’ correction and FDR p-value adjustment, comparing prevalence of non-zero dyad values in users. The second and third column report the percentage of users in the respective groups showing presence of dyads features. The fourth and fifth column report the number of users with non-zero dyad values. Dyads are ordered by difference in means, as in Fig. 3 of the main article. Rows where p-value $<0.05$ are in bold. The table shows that dyads with high average values in mainstream-prone users are employed significantly more often by those users, while there is no significant difference in the other dyads presence.

Dyad	MAp $\ne \mathbf{0}$%	MIp $\ne \mathbf{0}$%	MAp $\ne \mathbf{0}$ #	MIp $\ne \mathbf{0}$ #	${\text{χ}}^{\mathbf{2}}$	p-value
frozenness	53.4%	56.2%	13,223	230	1.154	0.377
submission	39.4%	39.1%	9,743	160	0.003	0.993
dominance	23.6%	26.2%	5,834	107	1.349	0.362
shame	45.4%	44.2%	11,235	181	0.171	0.761
ambivalence	20.0%	18.6%	4,962	76	0.456	0.608
aggressiveness	23.3%	23.2%	5,771	95	0.000	1.000
contempt	48.8%	45.5%	12,076	186	1.652	0.309
alarm	25.3%	23.9%	6,272	98	0.339	0.654
unbelief	14.7%	14.1%	3,632	58	0.045	0.897
sentimentality	41.1%	36.0%	10,189	147	4.344	0.069
outrage	22.0%	19.6%	5,440	80	1.244	0.371
pride	14.7%	13.0%	3,636	53	0.835	0.459
hope	65.2%	59.7%	16,123	244	5.119	0.051
envy	54.1%	48.7%	13,377	199	4.519	0.067
guilt	22.6%	18.6%	5,571	76	3.352	0.117
confusion	48.3%	42.1%	11,946	172	5.997	0.040
anxiety	37.6%	32.0%	9,312	131	5.153	0.051
morbidness	16.2%	13.2%	4,016	54	2.500	0.188
cynism	17.2%	12.2%	4,256	50	6.673	0.034
despair	65.8%	59.9%	16,290	245	6.025	0.040
disappointment	30.6%	24.2%	7,564	99	7.394	0.026
bittersweetness	31.6%	24.2%	7,820	99	9.869	0.008
remorse	53.5%	45.5%	13,238	186	10.087	0.008
love	70.6%	65.0%	17,459	266	5.633	0.045
pessimism	38.9%	29.3%	9,627	120	15.114	0.001
curiosity	42.4%	32.2%	10,504	132	16.660	0.001
optimism	61.4%	52.8%	15,205	216	12.293	0.003
delight	48.2%	39.1%	11,930	160	12.968	0.003

DOI: 10.7717/peerj-cs.3398/table-7

Addressing RQ3, the fear + anger dyad (frozenness) and the trust + fear (submission) dyads are those with the largest difference of mean values in favor of misinformation-prone users, while optimism and delight are those with the largest difference in favor of mainstream-prone users. These results confirm the hypothesis of an interplay between negative and positive emotions in misinformation-prone users, as explored in RQ2.

Emotional arborescence

Finally, we use the emotional arborescences introduced in subsection Emotional Arborescence of the Methods section, to analyze the structure of emotional co-occurrences in comments by misinformation-prone and mainstream-prone users. Figure 4 visually represents how a key/root emotion co-occurs with other emotions, with these latter emotions branching out from the root. The thickness of the stems and branches indicates the probability of co-occurrence, thereby highlighting stronger or weaker associations. Overall, the visualization resembles a tree structure with a root and multiple branches, though cycles between branches may also emerge. We examine the set of comments from misinformation-prone users and mainstream-prone users, focusing on the emotions expressed in these comments. We compute the probability $P(v_m,v_n|v_e,p)$ of finding emotions $m$ and $n$ in a comment containing emotion $e$ from a user with commenting preference $p$. Additionally, we perform an undersampling test on emotional stems (ES) and emotional branches (EB) of mainstream-prone user comments to ensure that any observed differences in emotional distribution are not simply due to the smaller number of comments from misinformation-prone users. Details on significance computation and robustness check are reported in the Methods section. Comparing the emotional arborescence between the two groups, we find that comments from misinformation-prone users exhibit a sparser emotional landscape, as depicted in Fig. 4. We then quantify the variety of emotional features in the two user groups by comparing the emotional arborescence of the comments from mainstream-prone users, ${\mathcal{A}}_{MAp}$, and from misinformation-prone users ${\mathcal{A}}_{MIp}$. By examining the ratio $\frac{{\mathcal{A}}_{MAp}-{\mathcal{A}}_{MIp}}{{\mathcal{A}}_{MIp}}$, we find that mainstream-prone users show $25\mathrm{\%}$ more emotional arborescence than misinformation-prone users, which addresses RQ4. These results further stress the patters highlighted by RQ2: the emotional contents of user-generated content written by misinformation-prone users offers less variety than the content from mainstream-prone users.

Discussion

In this study, we investigate the emotional content of toxic language produced by two opposing online communities: misinformation-prone and mainstream-prone users. The online debate is driven by an emotionally and politically polarizing topic–COVID-19. Adopting the EmoAtlas’ computational framework of Plutchik’s emotions (Semeraro et al., 2025), we evaluate the presence and co-occurrence of basic and nuanced emotions among users predominantly engaging with misinformation or mainstream content at increasing levels of toxicity.

While prior studies found both higher levels of anger in misinformation contents (Han, Cha & Lee, 2020) and similar levels of toxic language in comments on both reliable and questionable YouTube sources (Cinelli et al., 2021), our analysis shows that, although anger is overall less frequent in comments to reliable information (−2.54%), its prevalence specifically increases within toxic comments (+1.93%), in particular in Inappropriate and Offensive comments. In the other group (comments to unreliable sources) we observe alongside the increased negative emotion load a “normalization” of anger with respect to the toxicity found in comments. In other terms, we find that anger is more likely to appear using appropriate language in comments to unreliable sources rather than in comments to reliable ones. This quantitative finding importantly aligns with past findings about fake news diffusion among users of different political alignments (Ognyanova et al., 2020). Anger is also more likely to appear in comments employing violent language, which, however, are numerically much smaller in sample size. This re-contextualizes the usage of anger in reliable and unreliable information sources, where comments to reliable sources more likely use Inappropriate and Offensive language with anger. Conversely, comments to unreliable sources use anger in Appropriate and Violent comments. As reported also in Plutchik’s theory, anger is an emotional reaction to fight against external threats (Plutchik & Kellerman, 2013; Plutchik, 2001), hence it is expected that users express anger more frequently when using very high levels of toxicity and aggressiveness.

Also, misinformation-prone users tend to express language associated with trust and fear, often in combination with anger. To quantify this emotional complexity, we introduce the concept of emotional arborescence, measuring the emotional variety present in comments from each user group. We observe that misinformation-prone users generally exhibit lower emotional variety compared to mainstream-prone users.

Anger and fear represent two antagonistic responses to external threats: anger motivates aggressive confrontation, whereas fear triggers avoidance or flight behaviors (Plutchik, 1982; Bushman, 2002). Our findings suggest that mainstream-prone users display more aggressive and confrontational language than misinformation-prone users. However, interpretations of these emotional differences must consider the inherent limitations of online comments, as these alone cannot sufficiently determine users’ intentions.

Employing explainable AI techniques based on SHAP scores (Lundberg et al., 2020), we identify trust and fear as particularly distinguishing emotions among misinformation-prone users. Specifically, higher levels of trust and fear significantly contribute to classifying user-generated content as misinformation-prone. Our results thus underscore trust and fear as defining emotional characteristics within the comments of misinformation-prone users in our dataset.

Interpreting the higher prevalence of trust among misinformation-prone users requires considering the dual nature of trust (Castelfranchi & Falcone, 2010). Trust can arise either from cognitive reasoning—such as relying on a source perceived as credible—or from affective processes, such as emotional bonds. Using online comments alone, we cannot clearly distinguish these cognitive and affective components, thus limiting our ability to pinpoint the exact origins of trust. However, the elevated presence of trust among misinformation-prone users likely reflects a combination of affective and cognitive factors. Trust may signify increased positively valenced social cohesion within an in-group (Devos, Spini & Schwartz, 2002; Rathje, Van Bavel & Van Der Linden, 2021), or could result from specific reasoning tied directly to the debated topic, COVID-19. Future research employing existing dictionaries that measure in-group cohesiveness in online environments (Rathje, Van Bavel & Van Der Linden, 2021) could better quantify and clarify the socio-affective dimensions underlying increased trust observed among misinformation-prone users.

The second major pattern identified by our analysis is the stronger presence of fear among misinformation-prone users. Fear is a negatively valenced emotion elicited by perceived external threats (Plutchik, 1982). Within this dataset, fear could relate to diverse concerns such as COVID-19 itself, lockdown measures (Stella, Restocchi & De Deyne, 2020), or disrupted social interactions due to the pandemic. Future research employing topic modeling (Alipour et al., 2024) or cognitive network science (Stella, 2020) could quantitatively investigate the specific sources of fear in online discussions. Beyond this open research avenue, our findings clearly show the prominent role of fear among misinformation-prone users. Extensive psychological literature primarily associates fear with avoidance-oriented responses like fleeing, loss of control, negativity, and heightened alertness (Plutchik, 2001; Coelho et al., 2020). This characterization differentiates misinformation-prone users from mainstream-prone users, who, as previously noted, express aggression more prominently.

Our study also introduces emotional arborescence as a novel complex-network measure designed to capture the richness and complexity of emotional co-occurrences in textual corpora. Arborescence quantifies how various emotions co-occur relative to a reference “root” emotion (see Methods), resulting in an interpretable and robust measure of emotional richness, available through simple counting operations, cf. the EmoAtlas computational framework (Semeraro et al., 2025), rather than via black-box estimates available from large language models. Arborescence makes it easier to investigate also patterns among more emotions co-occurring together. For example, higher arborescence relative to trust implies that emotions other than trust co-occur more frequently with each other. Thus, arborescence can enhance understanding of how emotions such as fear and trust are integrated within online content produced by different user groups. Our analysis reveals lower emotional arborescence in misinformation-prone users, whose comments frequently emphasize trust and fear while exhibiting limited emotional diversity. Conversely, mainstream-prone users display richer interconnections among emotions. This difference might be due to psychological phenomena limiting the emotional expressiveness of misinformation-prone users, like rumination (Bushman, 2002), i.e., a tendency to circle around, think and express the same ideas. This interpretation calls for future research directions testing psychological phenomena in relation with misinformation consumption.

Recognizing the limitations inherent in our analysis can guide future research. Online platforms introduce contextual variability, as online interactions substantially differ across platforms and from offline communication. Future research guided by cognitive psychology (Stella, 2022) should investigate, in controlled laboratory settings, the extent to which System 2 (analytical and deliberate) cognitive processes might be impaired among misinformation-prone users. Moreover, it is important to clarify that our classification of users as misinformation- or mainstream-prone is based exclusively on their activity within specific content channels, without necessarily implying endorsement or adversarial behavior (Pennycook & Rand, 2020, 2021).

Despite these limitations, our findings provide quantitative evidence of emotional distinctions in responses to content from reliable vs. questionable sources, both at aggregate and individual user levels. These emotional differences offer valuable psychological insights and enhance understanding of online misinformation engagement. From this perspective, our results represent an exploratory yet innovative approach to investigating cognitive and emotional differences between misinformation- and mainstream-prone users on a large scale.

Supplemental Information