Sharing, commenting, and reacting to Danish misinformation: A case study of cognitive attraction on Facebook
Data publikacji: 12 kwi 2025
Zakres stron: 55 - 75
DOI: https://doi.org/10.2478/nor-2025-0003
Słowa kluczowe
© 2025 Petra de Place Bak et al., published by Sciendo
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Social media platforms play a prominent role in modern communication, with adults now averaging nearly 2.5 hours of daily use (Statista, 2024). Among the changes brought about by the digital transformation of communication technologies are the removal of editorial filters and the extension of the potential reach of the users’ content. This has resulted in an information abundance as various previously separated social contexts have come together (Marwick & boyd, 2011). Societal stakeholders, such as the World Health Organization, have expressed concern about the quality of information that might make reliable information harder to identify (WHO, 2020). Since the widespread adoption of social media platforms, digital misinformation has received increased mainstream and scholarly attention in recent years (for reviews, see Abu Arqoub et al., 2020; Bak et al., 2022).
Recommender algorithms play a key role in organising the vast amount of information on social media, but users’ engagement behaviours – such as what they react to, comment on, and share – also significantly influence content visibility. This raises the question of which factors determine what attracts attention and drives engagement. Acerbi (2019) suggested that misinformation spreads on social media because it aligns with certain human cognitive preferences. Building on this idea, we explore the literature on cultural evolution and content biases in human communication. In this context, content biases refer to content types that are more attention-grabbing, memorable, and/or likely to be shared (Stubbersfield, 2022). We aim to test the hypothesis that information that taps into the cognitive preferences for certain types of content can boost online engagement.
The continuous registration of user data on social media makes it possible to test the relationship between such factors of attraction and the decision to transmit or engage with information with high ecological validity. So far, studies on content biases in online information-sharing have primarily focused on the relationship between sentiment (i.e., negativity and positivity bias) and the number of shares (see, e.g., De León & Trilling, 2021; Schöne et al., 2021; Youngblood et al., 2023). However, earlier experimental studies have found evidence of a bias toward social information (e.g., McGuigan & Cubillo, 2013; Stubbersfield et al., 2015) and threat-related information (e.g., Blaine & Boyer, 2018; Nairne et al., 2007), among others. While a few studies have examined the frequency of factors of attraction beyond sentiment in popular content such as urban legends (Stubbersfield et al., 2017) and online misinformation (Acerbi, 2019), these studies did not evaluate the relationship between the presence of factors of attraction and engagement scores (i.e., shares, reactions, and comments).
Based on the literature on content biases, we chose to focus on the following factors: positivity, negativity, social information (i.e., information about third-party interactions), threat-related information, and intergroup-related information (for reviews, see Stubbersfield, 2022, 2025). Empirically, we drew on the case of misinformation because it is societally relevant and theoretically significant since misinformation spreads despite being maladaptive and potentially harmful, for example, if users adopt the wrong health behaviours. Further, misinformation may benefit from being malleable to be optimally attractive, whereas authentic information is limited to the truth. We addressed the following research questions:
RQ1. To what extent are factors of attraction (i.e., negativity, positivity, social information, and threat-related and intergroup-related information) present in misleading Danish Facebook posts? RQ2. How does the presence of factors of attraction and visual material influence the success of misleading Danish Facebook posts?
We analysed a dataset of Danish Facebook posts gathered from the Danish fact-checking agency TjekDet’s website. A considerable number of posts are associated with the Covid-19 pandemic, and therefore we chose to test the potential influence of this topic on engagement. Finally, since the layout of Facebook posts vary greatly depending on whether they include visual material, we also tested the effect of including an image or video.
In the study, we drew on the empirical case of misleading Facebook posts in the Danish context. In line with earlier work, we embraced an empirically grounded interpretation of misinformation, characterising misinformation as content identified as “partly or entirely false” by accredited fact-checking organisations (Nissen et al., 2022). Empirical contributions to understand and classify misinformation have primarily focused on English data from Twitter (Bak et al., 2022). On Twitter, a case study from the 2019 Danish national election suggested that misinformation is not common in the Danish context compared with the UK (Derczynski et al., 2019). However, compared with Twitter, Facebook is a much more commonly used social media platform, both globally (Statista, 2023) and in Denmark (DR Analyse, 2024). In 2023, six out of ten adults (15–75 years old) used Facebook daily, with some variations between age groups, according to Danmarks Radio’s media development report (DR Analyse, 2024).
Like other Nordic countries, Denmark has a high Internet penetration rate of 98 per cent, according to the
Research has shown that certain types of information are more likely to be noticed, memorised, and retransmitted than others (Stubbersfield, 2025), so-called transmission biases. These can be either content biases (e.g., negativity bias) or context biases (e.g., learning from prestigious or successful individuals) (Kendal et al., 2018). Apart from biases toward emotion-expressing content, empirical contributions to the study of content biases in naturally occurring human communication have primarily focused on their frequency in popular content types; for example, Stubbersfield and colleagues (2017) found factors of attraction in 92 per cent of a sample of popular urban legends, primarily emotional content. Similarly, Acerbi (2019) found that 86 per cent of a sample of fake news headlines contained elements that could be traced back to factors of attraction. In line with these findings, we expect the factors of attraction to be present in a large proportion of the misleading Facebook posts.
There is a growing number of studies that have modelled the relationship between the expression of sentiment or discrete emotions and online sharing, and a limited number of studies that have looked at intergroup-related biases (e.g., Rathje et al., 2024). To form testable hypotheses, we reviewed the main literature in relation to the factors included in the study – that is, negativity, positivity, social information, threat-related information, intergroup-related information, and visual material.
Negativity bias is well-documented within psychology (Rozin & Royzman, 2001). It has an intuitive evolutionary rationale: It is better to expect the worst and be positively surprised than the other way around. This way, sensitivity to negative information may increase the chance of survival. Transmission chain studies have revealed negativity to be more transmitted than positivity (Bebbington et al., 2017) and that ambiguous stories are rendered more negatively over repeated transmissions. An empirical study of two centuries of Anglophone fiction found a general decline in the use of emotional words, driven by a decrease in positive language (Morin & Acerbi, 2017).
Berger and Milkman (2010) were the first to test the effect of sentiment and emotion expressions on the sharing of content in digital media. Drawing on a sample of H1: Both positive and negative information will receive more engagement than neutral posts because it captures attention better, hereby increasing the chances of engagement.
Another content type that has received attention in transmission chain experiments is social information, such as gossip and information about interactions between third parties (e.g., McGuigan & Cubillo, 2013; Stubbersfield et al., 2015). In freely forming conversation, social information makes up approximately two-thirds of the topics discussed (Dunbar, 2004). The social brain hypothesis holds that the human brain evolved to live in complex social groups (Dunbar, 2009), for example, to support corporation and communication. Experimental findings suggest that both children (McGuigan & Cubillo, 2013) and adult (Mesoudi et al., 2006) participants are more likely to transmit social than non-social information. In addition, a study on urban legends found that social information is better recalled than non-social information (Stubbersfield et al., 2015). Finally, in the digital context, Acerbi (2019) found social information and celebrity information to be the most frequently present factors of attraction in a dataset of popular fake news headlines. Similarly, Berriche and Altay (2020) found half of their sample of Facebook posts to be about social information, but social information did not increase engagement. Instead, the main driver of engagement was phatic posts – that is, posts aimed at establishing a social connection. Based on the literature, we pose our second hypothesis:
H2: Social information will receive more engagement than non-social information, in line with the social brain hypothesis and because social media platforms are primarily used for exchanging social information.
Threat-related information concerns anything related to human survival, such as well-being and health. In an experimental study of the transmission of rumours, Blaine and Boyer (2018) found that threat-related information is transmitted more often than emotion-expressing content, even for threats that are very unlikely. Nairne and colleagues (2007) found survival-related material to be better remembered compared to control conditions, indicating that human memory has evolved to remember information relevant for survival. Similarly, Stubbersfield and colleagues (2015) found threat-related information to be better recalled than control material, and the combination of social and threat-related information to be significantly stronger than threat alone. A motivation to transmit threat-related information is that sources of threat-related information are evaluated as more competent (Boyer & Parren, 2015). Moreover, on Twitter, Mousavi and colleagues (2022) found that threats “to personal or societal core values of target audiences” increased the engagement. However, whether this generalises to a broader definition of threat-related information is currently unknown. We thus pose our third hypothesis:
H3: Threat-related information will increase the overall engagement score because it is adaptive to pay attention to threats and, being the source of threat-related information, can be associated with reputational benefits.
We also focused on intergroup-related information. Intergroup-related information differs from social information in that it describes relationships at the group level, for example, religious or ethnic groups, rather than information about identifiable individuals. It taps into intergroup bias, that is, the tendency to favour the group one identifies with (ingroup) over the outgroup (Hewstone et al., 2002). On Twitter, users tend to speak positively about their political ingroup and negatively about their outgroup (Stieglitz & Dang-Xuan, 2013). It remains unclear how intergroup-related information influences spread on social media, but information that aligns with preexisting social categories is selected for (Lyons & Kashima, 2001; Martin et al., 2017). Rathje and colleagues (2024) found posts about a political outgroup to be a stronger positive predictor of shares on Facebook and Twitter than the more established predictors of negativity and moral-emotional language. Our fourth hypothesis is thus as follows:
H4: Intergroup-related information will lead to higher engagement scores because it serves as a strategy for signaling group membership.
Finally, we consider the inclusion of visual material. The cognitive system has a longer history of engaging with visual material rather than text, for example, when our ancestors scanned their ecological surroundings. Visual material also differs from text because it stands out by being colourful and, in the case of video, includes motion, which might be efficient attention-grabbers in a crowded information space. Research from the field of advertising suggest that ads with visual elements grab attention better than purely textual ads (Goodrich, 2011; Pieters & Wedel, 2004). On Twitter, Bruni and colleagues (2012) found tweets with links to videos or images to be more retweeted than posts without multimedia content. In a more recent study, Li and Xie (2020) found that including an image in a tweet increased engagement. We thus propose our fifth hypothesis:
H5: The inclusion of visual material, that is, images and videos, leads to higher engagement scores because it grabs attention better than purely text-based posts.
The dataset was collected via the EFSCN and IFCN-certified fact-checking organisation TjekDet’s [CheckIt] website, where the fact-checkers exemplify claims rated as “entirely or partly false” through direct links to Facebook. Fact-checking agencies are a common data source in studies of misinformation due to the convenience of an already annotated sample (López-García et al., 2021; Song et al., 2021; Vosoughi et al., 2018). The data were collected by scraping all external links with the word “Facebook” from the “entirely or partly false” web page.
The described data-sampling method resulted in a list of links that were then manually visited by the first author to extract data about the posts to be used in the analysis. For each post, we collected the text, visual material, and the year the post was published on Facebook, as well as the engagement score (i.e., total number of reactions, shares, and comments). This approach resulted in a dataset of 356 Facebook posts, after four were excluded because they appeared more than once. The misleading claims were posted on Facebook during 2011–2023, with most entries being from 2021 (110 posts) and 2022 (95 posts). The dataset is considered sufficient in size for the purpose of this study (for similar samples, see Acerbi, 2019; Berriche & Altay, 2020), as the study relies on the idea that individual-level interactions result in population-level aggregates (Acerbi, 2016).
We chose engagement scores as an operational measure of online information sharing. On Facebook, users engage with posts through comments, shares, and reaction functions, such as “like”, “love”, and “wow”. Our analysis scrutinises these engagement types collectively and individually, exploring potential disparities in user behaviours to enrich our understanding of the dynamics surrounding misinformation dissemination. Engagement scores are well-established as measures of spread in the context of social media, especially in studies quantifying the effect of sentiment expressions (e.g., Berriche & Altay, 2020; Ferrara & Yang, 2015; Fine & Hunt, 2023; Schöne et al., 2021).
The data was manually coded for the presence or absence of the content factors. The posts were rated as positive, negative, or neither, thus treating positive and negative sentiment as mutually exclusive categories. Hence, each post could contain a maximum of four factors of attraction. The categories are defined according to the descriptions below (detailed instruction for coders can be found in Supplementary Material A).
The first author and a trained assistant independently coded 290 posts based on the absence or presence of factors of attraction from the five categories detailed in Table 1. To assess the reliability of the coding, the agreement was measured using Cohen’s kappa (
Categories and descriptions
| Social | The content concerns intense and noticeable social relationships (e.g., gossip, cheating, group alliances, controversies) but also everyday interactions and relationships (e.g., friends, family). |
| Threat-related | The content concerns possible threats (e.g., illness, violent acts, dangerous situations, death, and unvoluntary abortion). |
| Positive | The content expresses emotions that can be considered overall as positive (e.g., amusement, love, joy, etc.). |
| Negative | The content expresses emotions that can be considered overall as negative (e.g., sadness, regret, fear, anger, etc.). |
| Intergroup-related | The content concerns intergroup relationships, differences, or attitudes based on group membership (e.g., based in cultural, religious, social, and value differences). |
Concerning data access, the annotations and engagement scores are made available (OSF, 2023) together with the code used for scraping the external links from TjekDet and implementing the statistical analysis described below. This allows for repetition of the statistical analyses. To secure GDPR compliance, the Facebook posts cannot be shared. To convey the idea of how the posts look and were analysed, we have created the following mock example with an image generated by ChatGPT Plus:
Mock example of how the posts might look
This post would be rated as negative in tone due to the words “scam” and “lying” in combination with double use of exclamation marks. It also contains intergroup-related information, as the “politicians” are mentioned as lying and implicitly also responsible for the “scam”. It does not contain any explicit threats or social information, nor does it convey positive sentiment.
We evaluated the relationship between the coded predictors and the level of engagement by fitting a Bayesian zero-inflated negative binomial regression model. All predictors were modelled as fixed effects. Given the exploratory nature of the study, we chose conservative priors for our models that would constrain the models within a reasonable range (e.g., the models will not predict negative numbers of shares), without exerting any undue influence on the results. For the model intercept, we used a weakly informative prior, a student
All analyses were performed in R (R Core Team, 2023). Data were modelled using the brms package (Bürkner, 2017). Graphical posterior predictive checks were performed using the bayesplot package (Gabry & Mahr, 2022), model fit was further evaluated with the DHARMa package (Hartig, 2024), and Bayesian
Of the 356 Facebook posts, the majority (326) contained one or more cognitive attractive factors (92%). Of these, 114 (32%) contained two cognitive attractive factors, 101 (28%) contained three, 83 (23%) contained one, and 28 contained four (8%). 30 (8%) contained none
Figure 2 shows a difference at the topic level the proportion of posts containing each of the cognitive factors of attraction when the information is related to Covid-19 or not. In general, threat-related information, negativity, and intergroup-related information are the most frequently occurring factors of attraction. Intuitively, threat-related information and negativity are more present in the posts related to Covid-19 than those that are not.
Proportion of factors by Covid status (per cent)
Figure 3 shows the co-occurrences between the different factors of attraction. Most notably, social information often co-occurs with threat-related information (69%). Similarly, intergroup-related information often co-occurs with threats (79%) and negative information (64%). Threats are often formulated in a negative tone (58%).
Co-occurrences of factors of attraction (row baseline, per cent)
The modelled predictors accounted for 6 per cent of the variance in engagement (Bayesian
Incidence rate ratios of predictor variables on total engagement
Looking at Figure 4, the overall finding is that visual material (i.e., video and images) is the strongest predictor of increased engagement. Further, positivity, negativity, and intergroup-related information is also associated with increased engagement. Social information has a negative effect, and threat-related information, Covid-19, and the inclusion of emojis also reduce the overall engagement.
Table 2 shows the incidence rate ratios of total engagement and the three engagement types – comments, shares, and reactions – individually. The most noticeable aspect across engagement types is that the presence of videos is a far stronger predictor (numbers rounded) of shares (× 52.52) than comments (× 3.19) and reactions (× 3.02). The same pattern, although less pronounced, is shown for the other visual element, images, which increases the number of predicted shares by 7 times, comments by 2 times, and reactions by 3 times. Secondly, negatively valenced information influences shares (× 4.46) much more than comments (× 2.28) and reactions (× 1.65). Further, intergroup-related information resulted in 3.86 times the number of shares and 2.54 times the number of reactions but had no influence on comments (× 0.99). The remaining predictor variables show variance across forms of engagement.
Incidence rate ratios for total and subtypes of engagement (i.e., comments, shares, and reactions)
| Video | 21.66 (14.07) | 3.19 (1.91) | 52.52 (43.2) | 3.02 (1.69) |
| Image | 8.25 (3.18) | 2.8 (1.25) | 12.54 (7.44) | 3.72 (1.44) |
| Positivity | 3.36 (2.4) | 1.85 (1.34) | 2.76 (2.85) | 2.51 (1.6) |
| Negativity | 2.98 (1.21) | 2.28 (0.92) | 4.46 (2.32) | 1.65 (0.6) |
| Intergroup info | 2.73 (1.06) | 0.99 (0.39) | 3.86 (1.86) | 2.54 (0.88) |
| Social info | 0.88 (0.41) | 1.08 (0.48) | 0.92 (0.55) | 0.96 (0.37) |
| Threat info | 0.57 (0.26) | 0.82 (0.42) | 0.87 (0.45) | 0.65 (0.27) |
| Covid | 0.4 (0.16) | 0.15 (0.07) | 0.38 (0.2) | 0.36 (0.13) |
| Emojis | 0.11 (0.04) | 0.77 (0.3) | 0.05 (0.02) | 0.22 (0.07) |
We tested the influence of factors of attraction and visual material on the overall number of engagements (shares, reactions, and comments) that Danish misinformation received. The study showed a high frequency of the factors of attraction in the sample of misleading posts, with 92 per cent containing one or more factors. The high proportion of content that taps into content biases align with previous findings (e.g., Acerbi, 2019; Berriche & Altay, 2020; Stubbersfield et al., 2017). Hypotheses 1, 4, and 5 were all supported, as visual material (video and images), sentiment-expressing, and intergroup-related information received higher engagement levels. In contrast, hypotheses 2 and 3 were not supported, as social and threat-related information did not lead to higher engagement. This indicates that a high frequency of factors of attraction does not necessarily translate into high engagement.
Consistent with prior research (e.g., Bruni et al., 2012; Goodrich, 2011; Li & Xie, 2020; Pieters & Wedel, 2004), our findings indicate that visual material results in higher engagement levels compared with non-visual content. This outcome is expected, given that visual content is particularly effective in capturing attention amidst the vast array of information on social media. Importantly, this effect appears to be amplified in the context of misinformation. Visual material can be perceived as corroborative evidence, thereby reducing the perceived reputational risks associated with sharing potentially false claims (Altay et al., 2022). In essence, visuals may serve as a heuristic for validation, encouraging users to share content they might otherwise hesitate to disseminate. An example from the data is posts sharing the trailer for the conspiracy theorist “documentary”,
That threat-related and social information decreases the engagement level can be interpreted as an indication that these biases are primarily occurring in the memorisation and recall phase of transmission. Recall is not a requisite for online information-sharing where users can simply use the share function. This interpretation aligns with Acerbi’s (2021) experimental findings, in which the threat-related information was favoured only in the condition of mimicking oral transmission and not in the digital-sharing condition. Further, experimental work on urban legends has found social and threat-related information to be advantageous in the encode-and-retrieve phase (Stubbersfield et al., 2015), but not in the choose-to-receive and choose-to-transmit phases. The finding that threat-related information decreases the engagement score aligns with previous findings on Facebook (Berriche & Altay, 2020). A viable interpretation would be that the cognitive constraints on memorisation and reproduction in oral communication are what causes the bias toward threat-related information. Alternatively, or perhaps additionally, Covid-19 was a prominent theme in the data, which might have caused users to experience threat fatigue.
The affordances of social media – for example, that what we engage with is visible and persists across time and space (Bucher & Helmond, 2018) – may further moderate what we engage with. In the case of social information, it is not associated with an increase in engagement scores; however, it might still be attention-grabbing. Instead, users may abstain from engaging with it because of reputational concerns, such as not wanting to be caught gossiping and, in response, reserve such information for more private settings. Intergroup-related information, on the other hand, is less risky to share because it signals group membership and mitigates the risk of being caught gossiping. Future research could compare viewership data to engagement scores to test whether there are some types of content that have high numbers of views but little engagement. Alternatively, misleading social information might not spread because the wealth of authentic social information available online overshadows such content in terms of salience. Finally, traditional social media platforms such as Facebook rely on social connection (Metzler & Garcia, 2023), in contrast to newer, more heavily algorithmic-driven platforms such as TikTok. It is possible that the goal of maintaining social relationships overrides the biases toward social and threat-related information. As for positivity and negativity bias, large-scale studies have found that expressions of valenced sentiment increase spread on social media consistently across several languages (e.g., Mousavi et al., 2022; Tsugawa & Ohsaki, 2015). This aligns well with previous findings that emotion is privileged in early attention capture (Brady et al., 2020). Thus, our study indicates that these previous findings also generalise to the case of Danish misinformation.
Finally, the analysis shows that the three most frequent factors were threat-related information, negatively valenced information, and intergroup-related information. A plausible reason why these three biases are dominant is that Covid-19 was a prominent theme in the dataset. A high proportion of the dataset conveyed threats, for example, warning against vaccination, and sought to attribute responsibility for the pandemic to the outgroup, for example, the political elite or ethnic minorities. However, the study also indicated that there is not necessarily a relationship between the frequency of a trait and its transmission via engagement. That said, frequency does signal that these are factors that misinformed users in their communication on social media. Thus, frequency can be interpreted as a different signal indicating a preference in the choose-to-transmit phase, in addition to engagement scores. Further, the discrepancy between the prevalence of certain factors and their impact on engagement can be interpreted in multiple ways. Users might be unaware that visual material, for instance, significantly boosts engagement. Alternatively, it may reflect the higher costs associated with producing visual content, such as time and technical skills. Additionally, users may have objectives beyond maximising engagement. For instance, they might prioritise sharing threat-related information quickly to alert their network and to be perceived as cooperative and valuable members to the group.
Factors of attraction influence interactions, comments, and shares differently. Analysing the engagement types individually (see Table 2) reveals that the primary effects are driven by changes in the engagement type shares, especially for visual material. The share category exhibits the greatest variance, with a median of 3 and a maximum of 406,000 shares for a single post (see Supplementary Material C, Table C2). A plausible explanation for why visual material and factors of attraction primarily influence sharing behaviour is that users may be reluctant to share information publicly if they doubt its veracity. However, when a post includes visual content, it may serve as additional support for the claim, making users more likely to consult others about the veracity of the claim. In such cases, the share function offers a strategic advantage, allowing users to consult trusted others by sharing in private chats, closed groups, or on their own profile with restricted audience settings. In contrast, reacting or commenting on a post immediately exposes a user’s stance – such as showing anger or agreement – to a large and undefined audience, and thus, potentially involves reputational damage. This suggests that reactions and comments may be preferred when users are more certain of the claim’s truthfulness, either to support it or actively debunk it.
The observation that the types of engagement contribute differently to the overall score suggests that distinguishing different modes of engagement can provide insights into how various types of content appeal to specific behaviours. Compared to oral communication, the share function most closely resembles transmission. From a user perspective, reactions and comments are more subtle ways of extending a post’s reach. However, both comments and reactions still signal to other users and the algorithm that the post merits engagement.
According to the theory of cultural evolution, all other things being equal, cognitively appealing content will spread more than equivalent non-appealing content. However, such conditions are only obtainable in controlled experimental settings. On social media, algorithmic content curation, context biases, and social factors are also at play. This may explain why the model implemented in the study only accounted for 6.3 per cent of the variance in engagement. While this may appear small, considering the scale of social transmission on digital social media, 6.3 per cent can make a large difference in what users see. While it is beyond the scope of this article to address in detail the moderating factors introduced on social media, we briefly discuss them to illustrate that the processes of cultural attraction do not happen in a vacuum. When discussing cultural transmission on social media, it is relevant to consider a broader context; thus, this study feeds into a wider research agenda on why some information spreads while other information vanishes.
Algorithmic content curation may cause the over-dispersed tendencies of social media engagement scores (e.g., Apenteng et al., 2020; Gross & Von Wangenheim, 2022), with a few viral posts and many more that go without engagement. A possible assumption in this regard is that algorithms are optimised for prolonging attention, for instance, by amplifying the reach or visibility of negatively valenced information. In a similar vein, algorithmic meddling can also limit the reach of certain content. However, understanding how algorithms work remains a challenge. Additional factors that interfere with the results may be users’ perception of the audience and concerns for reputation management. In addition, context factors such as the number of followers or level of popularity may also influence what gets commented on, shared, and reacted to. That said, de Oliveira and Albuquerque (2021) found that social prestige does not determine spread of misinformation on Twitter. In short, the reasons for both sharing and engaging with online information are manifold, but cognitive factors of attraction seem to be part of the explanation.
The first limitation this study faces is the relatively small sample size. However, this was also what made manual coding feasible. Future studies can adapt the methodological design and scale it through natural language processing tools; however, this would require training the natural language processing tools to categorise other cognitive factors of attraction besides sentiment and emotions. Further, whether the observed patterns transfer to other social media platforms and cultural contexts is a task for future research to unveil. However, as discussed, the findings align well with previous efforts in different national contexts. Finally, the timespan of the data stretches across several years, and it is possible that changes to algorithmic content curation in the same period accounts for some of the difference in engagement rates. An obvious extension of the study would be to test the same hypotheses on authentic information to see whether the patterns transfer across types.
Moreover, the size and sampling strategy of the dataset impose some limitations on the generalisability of our findings. TjekDet selects posts for fact-checking based on their general relevance to the public (TjekDet, 2020), which may result in a dataset that is not fully representative of misinformation, and even less so of general information. Nevertheless, our primary objective was to expand on studies examining the frequency of factors of attraction in popular online content (Acerbi, 2019; Stubbersfield et al., 2017) by evaluating their relationship with engagement metrics. Future research could test the robustness of our findings in larger and more diverse datasets, potentially by utilising advancements in language models for more comprehensive annotation.
Finally, studying cognitive factors of attraction in an ecologically valid setting is messy, as multiple factors other than the included factors exert influence on the decision to engage with information, for example, reputational concerns (Altay et al., 2022), and the entanglement of algorithmic and social drivers (Metzler & Garcia, 2022).
Despite these limitations, this article presents an interdisciplinary methodological framework combining cultural attraction theory and media studies and a fine-grained analysis of how factors of attraction scale in contemporary media systems. Additionally, the article adds to the field of misinformation studies by focusing on a platform other than Twitter, in an understudied cultural context (Bak et al., 2022), and by elaborating on the cultural success of misinformation (see also Acerbi, 2019; Berriche & Altay, 2020; Youngblood et al., 2023).
In this article, we have shown a high frequency of factors of attraction in a sample of misleading Facebook posts. Still, only visual material, positivity, negativity, and intergroup-related information are associated with higher engagement scores. The effect is primarily driven by the engagement type share, indicating that the different types of engagement – shares, comments, and reactions – mean different things. Overall, the model can explain 6.3 per cent of the variance in engagement, suggesting that included factors of attraction do influence social media communication. However, the specific affordances of social media may moderate certain types of biases; for instance, intergroup-related information might be shared more than social information to avoid reputational damage from openly gossiping. Visual material has, unsurprisingly, the strongest influence on engagement scores, indicating that it is an efficient attention-grabber. A future step would be to augment the analysis to a larger dataset by training a classifier. This would allow us to determine whether the findings presented here generalise outside the context of misinformation on Danish Facebook.