Good and Bad Science? Spanish-Speaking YouTube Users’ Comments on Science Content

Abstract

In recent years, affective polarization has reached issues that were (at least previously) considered apolitical (see Rudolph & Hetherington. Affective polarization in political and nonpolitical settings. International Journal of Public Opinion Research, 33(3), 591–606, 2021. doi:10.1093/ijpor/edaa040). Perhaps the citizens’ reaction to coronavirus disease-2019 has helped to bring this relationship to its peak. This research proposes to analyze the response of YouTube users to the most popular videos on climate change, health, technology, and science in Spanish-speaking countries. To do this, the present work proposes an analysis using deep learning techniques. We selected the 50 videos with the highest number of views for each topic. Then, we adapt the linguistic models used to obtain the articles to index the degree of polarization. The strategy was twofold: first, using ratios and fluctuations between words; second, by measuring the affective distance both between the videos and the comments and between the comments prioritized by the platform’s moderation. The results show interesting data. First, the Spanish-speaking population does not behave similarly to the populations of Southern Europe, which are culturally related. Second, affective distance (which we propose is an indicator of polarization) does not grow over time and is not directly related to active participation in social media.

Introduction

In recent decades, contemporary societies have undergone significant changes due to rapid innovation and modernization. These transformations make citizens subject to different conditioning factors. According to Inglehart (1977), values have mutated in the Western world’s most prosperous regions since the second half of the last century. This silent revolution has generated a new way of understanding reality in the public, political, and social spheres (Inglehart & Welzel, 2005). These new cohorts, born as early as the late 1990s and known as digital natives or the net generation, have a special relationship with technology (Helsper & Eynon, 2013). Certainly, social scientists wonder about the potential effects of those changes, which led to a vivid debate. In this sense, it is essential to highlight the role that the media (and its potential consequences) have played in driving and establishing liberal regimes (Bennett & Entman, 2001; Gurevitch, Coleman, & Blumler, 2009), serving as both a watchdog and a counterweight (Trappel & Tomaz, 2021). With the massive arrival of television, the media overturned what had been laid out by the press and radio in previous decades. That created new cultural and social narratives that reached an ever-widening audience (Kellner, 2020). However, the emergence of the Internet changed the landscape again. Traditional media, once essential agents of socialization, began to lose prominence. Social media became the paradigm of the Internet. The integration of these platforms expanded the thesis of an interconnected world. They have revolutionized how we communicate from Facebook to Twitter, from YouTube to Instagram.

These processes have stimulated contemporary societies, moving from an analogue to a fully digital world rapidly, both in how we receive information and interact with others and communicate our opinions. Hence, since their emergence, these transformations have captured the attention of numerous scholars (Cunha, Alves, Araújo, & Ares, 2021). The rapid growth of social networks and the constant presence of multiple screens in our daily lives (Chadwick, O’Loughlin, & Vaccari, 2017) have generated diverse approaches from a psychological and medical perspective (Jung, Ju, Song, & Lee, 2016) and from their role in helping organize the social movements that have emerged since the beginning of the 21st century (Young, Selander, & Vaast, 2019).

New reflections emerge when considering social networks as new media, different from traditional media. Recent terms such as bubble filter or Google wall emerge (Dillahunt, Brooks, & Gulati, 2015; Michiels, Leysen, Smets, & Goethals, 2022; Pariser, 2011), while others, such as echo chambers, are revisited (Cinelli, De Francisci Morales, Galeazzi, Quattrociocchi, & Starnini, 2021; Guess, Nyhan, Lyons, & Reifler, 2018; Serrano-Contreras, 2021). Both have been part of new studies on radicalization, disaffection, or political tension. However, the common point of all these concepts is the search for an explanation for the apparent increase in affective polarization in many societies.

From distancing to polarization?

Over the past years, there has been a growing interest in political polarization (Hernández, Anduiza, & Rico, 2021), probably due to the rise of this phenomenon worldwide (Gidron, Adams, & Horne, 2019). This interest is primarily driven by the search for explanations for events that defy the expected logic of social action (Hameleers, 2019; Maher, Igou, & van Tilburg, 2018) concerning the success and gradual improvement of the quality of democracies (Carothers & O’Donohue, 2019; Duell & Valasek, 2019; Kalmoe & Mason, 2022). Unlike early studies on political polarization that focused primarily on the United States (Abramowitz & Saunders, 2008; Bafumi & Shapiro, 2009; Darmofal & Strickler, 2019; Iyengar, 2021; Levendusky, 2009), current contributions are broadening the focus. One meaningful change has been a redefinition of the term (Kekkonen & Ylä-Anttila, 2021; Wagner, 2021). When addressing the question of what political polarization is, one finds the basic concepts of two poles that are far apart on an axis (Bramson et al., 2017; Tucker et al., 2018). Extremes that spatially may also be, referencing Di Maggio, Evans, & Bryson (1996), distancing or already distancing. Moreover, as Bramson et al. (2017) point out, that distancing can be found within previously homogeneous ideologies.

Pioneering studies on political polarization were based on the spatial distribution of ideological positions (Mason, 2015). This type of polarization proved useful for modeling citizens’ and elites’ positions on political issues (Abramowitz & Saunders, 2008; McCarty, Poole, & Rosenthal, 2005), as the work of Poole & Rosenthal (1984, 1985) showed. However, considerations on the issue have looked for other features that can adapt to changes threatening social coexistence and well-being in liberal democracies (Boxell, Gentzkow, & Shapiro, 2020; Carothers & O’Donohue, 2019). We can find literature in public opinion describing the existence of different variables that operate as polarization moderators, that is, the context (Druckman, Klar, Krupnikov, Levendusky, & Ryan, 2021; Garzia, Ferreira da Silva, & Maye, 2023), partisanship (Druckman, Klar, Krupnikov, Levendusky, & Ryan, 2013), level of knowledge (Trilling, van Klingeren, & Tsfati, 2017), the degree of tolerance (Xia & Shen, 2023), internal efficacy (Stapleton & Wolak, 2024), or the exposure to traditional media (Kim, 2019), which in the latter case could be interesting regarding our proposal: science materials.

Cultural moderators appear to be secondary to those mentioned above, at least when approaching political polarization. However, works such as Fletcher and Jenkins (2019) expanded the scope and added conditions already exposed by the work of Hallin and Mancini (2004), who, correspondingly, also underlined the differences between European and Anglo-Saxon countries. Likewise, Reiljan (2020) finds persistent singularities in European and American countries and their party systems when correlating ideology and affectivity. Other studies, such as that of Luengo, García-Marín, and de-Blasio, (2021), expose the relationship between traditional media systems and the communicative spheres of social networks. It is plausible that those differences between cultures could very well have transferred from traditional to social media.

In this way, several approaches show how social networks can exacerbate the feeling of belonging to a community (Bessi et al., 2016), the formation of large clusters (Mentzer, Fallon, Prichard, & Yates, 2020), as well as the simplification of discourse or the use of aggressive or hateful language (Buder, Rabl, Feiks, Badermann, & Zurstiege, 2021; Marchal, 2022). In accordance with this, new research dynamics combined different disciplines, such as political science, sociology, psychology, and language processing (Serrano-Contreras, 2021; Serrano-Contreras, García-Marín, & Luengo, 2020), to measure these social interactions. These contributions shift the consistency of the ideological perspective (Lelkes, 2016) to focus on the analysis of the feeling of belonging (Huddy & Khatib, 2007; Levendusky, 2018): tribalism and affectivity toward one’s group (Suhay, 2015; Whitt et al., 2021); aversion (Mason, 2018); or dehumanization toward the other (Harel, Jameson, & Maoz, 2020). Thus, a number of scholars consider the Internet and social networks to be drivers of actual phenomena, although there are also cautious positions (Barberá, Jost, Nagler, Tucker, & Bonneau, 2015; Beam, Hutchens, & Hmielowski, 2018; Boxell, Gentzkow, & Shapiro, 2020), and even those that have identified an inverse process: increased polarization drives greater consumption of social networks (Nordbrandt, 2021), an idea also identified by Settle (2018) in Frenemies. Bail et al. (2018) highlight another paradox: the intensification of differences when we find ourselves in environments with opposing positions to our values and ideology. However, as highlighted by authors such as Serrano-Contreras, García-Marín, and Luengo, (2020) or Luengo, García-Marín, and de-Blasio, (2021), dynamics in social networks do not always reach extreme values with increased consumption or deliberation. They highlight that there are factors that moderate the results of polarization. For example, in the case of YouTube, the algorithm driving the consumption is not necessarily leading to radical content (García-Marín & Serrano-Contreras, 2023). The work of Fishkin, Siu, Diamond, and Bradburn, (2021) also supports the idea, although, in this case, it does not use computational models. In any case, it seems that deliberation is somehow related to polarization. Furthermore, some contributions imply the direction of that relation: intra-group deliberation can alleviate and serve as an antidote to extreme positions (Grönlund, Herne, & Setälä, 2015), generating consequences such as de-polarization (Xia & Shen, 2023). Then, we can assume that citizens’ activity on social media will be positively linked to polarization, measured by affective distance. Argyle and Pope (2022) also agree on this process when identifying political participation as a predictor of polarization. One of the dimensions chosen in that model is persuasion, which is undoubtedly also present within the activity in social media, which we interpret here as a minor sphere of participation. In our case, we equal participation to deliberation, using the number of comments as an indicator. Hence, in our first hypothesis, we state that the number of comments (activity on social media) will be positively related to affective distance.

Nevertheless, there are more moderators: time has been considered by several studies an essential variable to understand polarization, although we can find contrasting inferences. For example, Trilling, van Klingeren, and Tsfati (2017) found that ideologically congruent exposure (in a European multi-party setting) does not necessarily provoke polarization over time. However, we could also presume the opposite (probably most likely in bipartisan systems). Kim (2019) argues that exposure is not significantly linked to polarization in the mid-term and that general political knowledge somehow moderates polarization. Political tolerance (mainly present in scientific contents) could also work as an increasing factor or reduce polarization over time (Xia & Shen, 2023). However, some of these trends might have changed in recent years, especially with the politicization of crises such as coronavirus disease 2019 (COVID-19) (Kerr, Panagopoulos, & van der Linden, 2021). The global epidemic challenged political systems worldwide and led to heated debates in networks, where misinformation acquired a new dimension (Yang et al., 2021). One possible immediate effect was the politicization/radicalization of topics previously not the subject of social debates or, at least, not the main focus of public opinion scholars. Confidence in government measures to deal with the emergency was shaken (Fasce, Adrián-Ventura, Lewandowsky, & van der Linden, 2023) and linked to geographical and cultural variables (Klymak & Vlandas, 2022; McLamore et al., 2022). As Bergan, Lapinski, and Turner (2022) point out, expert rejections can engender the opposite effect on public opinion, generating context-dependent trust. The coronavirus crisis serves as a paradigm where cultural and pertaining factors intersect, even calling scientific issues into question. Thus, following Rekker (2021), beliefs or ideologies can be a transmission belt for rejecting scientific methods and advances. These issues reveal symmetrical group confirmation biases among both conservatives and liberals (Ditto et al., 2019), biases that are also found among researchers themselves (Lieberman & Long, 2022). That deliberation growth could translate into an increased polarization on health and science topics. Hence, our second hypothesis is that affective distance on science-related topics will increase over time.

Method and Data

It is clear, then, that there are multiple moderators in the relationship between social media use and polarization. The type of topic under discussion seems crucial (Serrano-Contreras, García-Marín, & Luengo, 2020), as well as social and cultural factors (Luengo, García-Marín, & de-Blasio, 2021), or the social network itself (Van Bavel, Rathje, Harris, Robertson, & Sternisko, 2021). This research focuses on a particular set of topics, such as those related to science and technology. Furthermore, it also focuses on a Spanish-speaking sample and a social media less commonly analyzed, that is, Youtube. Our starting hypotheses are:

  

Data Collection

An essential characteristic of this study is its focus on analyzing polarization beyond the North American or Western European contexts. Consequently, the samples have been selected to investigate the dynamics within Spanish-speaking regions. Although we do not correct the sample by localization, we understand that most users come from the Americas, reflecting the population imbalance (less than 10% of native Spanish speakers are in Europe). By doing so, we aim to provide a broader perspective. As stated above, we have chosen YouTube as the platform for the analysis. This social video network, owned by Alphabet, has experienced significant growth in recent years; YouTube was ranked in 2023 as the second among all global social media platforms, with around 2,500 million active users, and with more than half of its users born between 1981 and 2012 (Dean, 2024). Thus, our sample collection process involved several phases. Initially, we conducted a parameterized search on YouTube using four topics with Spanish keywords: “climate change,” “science,” “health,” and “technology” (“cambio climático,” “ciencia,” “salud,” and “tecnología”) (see Figure 1 for the evolution of the sample over time). Those themes also serve as sparks between groups, bringing us closer to the idea proposed by Rekker (2021). From that search, we selected the 50 most viewed videos with comments enabled and relevant to our research interests. This initial selection yielded 200 video links, forming our data corpus. Each topic mainly had videos directly related to the topics they referred to, but some videos were off-topic.

The videos had different authors: on “climate change,” 15 of the videos belonged to the German news channel DW, and 5 were from different media, while another 20 videos belonged to science communicators. The remaining (10) are divided between the Spanish company Acciona (2), documentaries (3), interviews (2), one about a song about the issue, one video from a YouTuber, and one from a Colombian university channel. As for the videos on “health,” 27 come from content creator channels, 5 from mainstream media, 2 from conferences, and another 2 from companies, in this case, a bank and an insurance company. Remarkably, 13 are videos related to Christian prayers and another video on mystical exercises. The “technology” videos comprise 26 videos from content creators, 12 from generic content, and 14 from scientific communicators. Another 21 videos come from different media outlets, although 12 are from the same source, TechZone. Finally, 2 of the videos are from company channels, and the last is a documentary. Finally, 45 videos are from content creators within the “science” theme, mainly focused on science communication. The remaining 5 videos are 3 from a bank, one from a Spanish university channel, and a final video on religious topics.

For our data collection, we employed the Python library YouTube-transcript-API (Depoix, 2018) to download the verbatim transcripts of the provided video IDs. Simultaneously, using the YouTube API, we obtained the video comments, tagging each one with its corresponding topic. This meticulous process allowed us to amass a sample of n = 769,546 comments, ensuring the comprehensiveness of our data.

Table 1 shows the number of comments by topic, with “science” followed by “health” being the most commented. Meanwhile, “technology” and “climate change” reached less than 200,000 comments.

Research Design

We propose investigating polarization using natural language processing techniques. We aim to index group distancing. In order to explore new perspectives in the study of polarization—as claimed by Kubin and Sikorski (2021) —we propose a formula for identifying factional separation. In particular, we focus on affective or semantic distancing, and we rely on the works of Serrano-Contreras, García-Marín, and Luengo, (2020), García-Marín (2021), or García-Marín and Serrano-Contreras (2023), which offer relevant contributions on this type of analysis. The novelty of this work is the incorporation of deep language computational models (Kaur & Sharma, 2023; Saxe, Nelli, & Summerfield, 2021), valid for automatically assigning scores to each of the units of analysis.

As shown in Figure 2, the first step in calculating the affective distance is a sentiment analysis. That analysis was conducted with sentiment.ai¹ library using the “multi.large”² model, adapted to languages other than English and with good accuracy in Spanish. Of course, we chose a small sample to check the codings (being the unit of analysis, a commentary or a complete video) without finding any inconsistencies (Ophir & Walter, 2023). Affective distancing involves calculating the average sentiment for each unit of analysis, a video or a comment, and calculating the distance between the sentiment analysis of the comment and the video (in absolute numbers).

That gives us a number that can take any value between 0 and 2, which is the difference in the sentiment of the comment relative to the video, but not the negativity or positivity of the comment. We understand that positive or negative comments may or may not be polarizing, but it always depends on the context. However, distance in sentiment can be an indicator of disagreement or polarization. That has been applied quite successfully in the research cited above.

The subsequent phase involved refining the sample to address imperfections and encoding errors. We used the R programming language and executed the necessary procedures to achieve this. It is important to note that, unlike traditional natural language processing approaches, we abstained from applying conventional preprocessing techniques such as removal of stopwords or elimination of punctuation marks. That was motivated by the understanding that deep learning models improve accuracy when provided with a broader range of information. For this purpose, the sentiment.ai library conducts a word embedding process without relying on a rigid lexicon. As described by the authors (Wiseman et al., 2022), this library employs a strategy akin to principal component analysis (PCA) models to generate information and scores. Finally, with the scores obtained (ranging from 1 for positive to −1 for negative), we constructed an index to calculate the affective distance. This measurement quantifies the ratio between the sentiment score of each video and the sentiment score of its corresponding comments, namely affective distance.

Results

To verify our hypotheses, we present different explanatory descriptions of the types of videos selected, as well as the values obtained by measuring the affective distance for the comments.

The results in Table 2 indicate that there could be a more pronounced affective distance between transcripts and comments—both their mean and median values and their standard deviation show that there are not very extreme positions. The most extreme value is 0.51, away from the maximum ratio.

Figures 3 and 4 show the first results of the comments. In the first case, we aggregate the results according to the density of the sample, together with the number of likes of the comments. It can be seen how the affective distance follows a nonlinear descending pattern. That means that most of the comments are on the same line as the videos (regarding the number of likes, we see that it behaves similarly).

The most interesting information is to check whether affective distance behaves similarly across the topics analyzed. Furthermore, Figure 4 shows substantial differences between them: We can establish two clear patterns. The first is constituted by “Technology” and “Climate_Change.” It is easy to appreciate that they share a similar behavior that could be due to a polarization of the comments: The comments are placed at two extremes, resulting in a U-shaped graph. That means that the comments are in two groups, one very close and one opposite to the positioning of the video. The phenomenon is not very obvious in the case of “Technology” but is very marked in “Climate_Change,” with a more uniform distribution in the values of the affective distance.

The other two themes, “Health” and “Science,” also show similarities. In both cases, there is a decrease in the number of comments as the affective distance increases, an apparent phenomenon with the first topic and much less so with the second. In aggregate terms by topic, Figure 5 shows how “Climate_Change” and “Technology” have higher polarization averages (and wider distribution) than “Science” and “Health.”

Given that the distribution is normal, a simple analysis of variance shows that the differences in means are significant (df = 3.0, sum_sq = 5,443.720191, mean_sq = 181,414.573397, F = 14,894.99 for p < .001).

We can also approach the results from a chronological point of view. As the sample has been selected without time criteria, only popularity, both the videos and the comments have dispersed over the years, in some cases since 2012. That allows us to see the behavior of the affective distance chronologically. Figure 6 shows this information. The first row indicates the aggregates of comments and the average behavior by month. The comments have grown, albeit with significant peaks, and the average affective distance has decreased. In general terms, this means that our H2 is negative. It is essential to highlight some relevant data in this figure, such as the periodic peaks in comments observed in the “Science” topic. One possible explanation is that these conversations are generated on videos related to the topic. This situation could influence the moderation of comments since, as the number increases, the average values tend to standardize. However, it is relevant to remember that a large part of these videos, precisely 45, is aimed at a specific audience interested in science popularization. In the Spanish-speaking world, there are several channels, such as the one analyzed, Quantum Fractum, which has many followers and constantly shares this type of content. The audience for these videos tends to show little distance, as many of their followers comment to express appreciation for the content shared without entering into debates or divergent comments.

The bottom row provides similar information but by topic. Here, we can observe essential differences in the distributions of the primary variable. First, the number of comments on “Science” presents significant peaks. The second and more evident is that the behavior of the affective distance tends to be stable over time, although the variations, especially in the case of “Climate_Change,” are very significant. That is probably because the sample is concentrated in recent years.

The values obtained in Table 3 are small because they show the monthly increments of a variable that can take values from 0 to 2. There are only two significant relationships between the time variable and affective distance: specifically only for the topic “Health” a significant inverse relationship is observed (y = −0.0028x + 0.8384, R2 = 0.802, p < .001); for “Technology,” the relationship is direct but less relevant (y = 0.0019x + 0.4104, R2 = 0.503, p < .001). To accurately assess the impact of time on comments, we have created a variable, “date_dif,” which measures the distance, in days, of the comment from the publication of the video. Subsequently, we created groups according to the distribution of these comments (Figure 7); as can be seen, in all the topics, most of the comments occurred between 15 and 30 days after the publication of the video and, from there, progressively towards the publication date.

How does affective distance relate to the comment’s publish date? Figure 8 shows that the differences change depending on the topic, supporting the idea that the topic strongly moderates user behaviour. Overall, the changes are minor in “Health” and “Climate_Change” (where there are slight increases) and more significant in “Technology” and “Science” where they tend to decrease over time.

Another question is related to users’ behavior toward comments, information that we can extract from the likes (whose density can be found in Figure 4 and has served as the basis for grouping them into the variable “Likes”). Here, the information obtained is more relevant as it implies action on the users’ part. Furthermore, it can be seen in Figure 9 how the number of Likes affects the affective distance.

The results are very interesting, showing how the number of Likes decreases as the affective distance increases. Alternatively, as shown in Figure 9, comments that are more affectively detached from the videos (arguably more polarizing) receive less user approval. These results are surprisingly inconsistent with previous findings; Luengo, García-Marín, & de-Blasio, (2021) show that British users tended to receive fewer likes when comments were more polarizing, while Italians and Spaniards behaved the opposite way. Our results seem to confirm that, at least in the cases analyzed, the wider Spanish-speaking public behaves more like the British than the Mediterranean Europeans. Therefore, as we thought, these results reinforce the idea that divergent contexts (and topics) show different outcomes when understanding polarization dynamics.

Furthermore, the results obtained from the non-parametric test (since “Likes” does not behave like a normal distribution) of comparison of means (Mann–Whitney U) between different topics (“Technology,” “Science,” “Health,” and “Climate Change”) reveal significant differences in terms of the number of likes. The values of the test statistics for all comparisons were relatively high, indicating substantial differences in the means of likes between the different topics. The p-values obtained were extremely low (all less than .05) for all comparisons made, indicating a very high statistical significance. That supports the conclusion that the means of likes differ significantly across topics, which supports the conclusion that user behavior is not homogeneous across topics. In particular, we observe that Climate Change shows the most pronounced difference compared to the other topics, with a p-value almost equal to zero. That suggests that the “Climate_Change” topic generates a unique response regarding likes compared to the other topics analyzed.

To demonstrate our H1, we tried to correlate the number of comments with the affective distance without success. However, to further explore this relationship, we proceeded to analyze comments with replies and comments that are replies. Table 4 is striking in that the comments with responses show a lower affective distance than the average for the topics, that is, they would be more in tune with the video. That again implies that comments with a greater affective distance would have fewer responses. However, also in all topics, the responses show a greater affective distance.

What does this data show? That people who reply to user comments do so in the opposite direction to the average number of comments on the video, but get fewer likes than the average number of comments. Therefore, the data suggest that greater polarization would not encourage more significant activity in social media but rather the opposite. For the sample analyzed, less polarization would encourage greater but more polarizing comments. As we have mentioned, this would not verify our second hypothesis, although we cannot rule out some relationship, albeit moderated by the comments to the videos. If we choose only the responses to other comments instead of the whole sample, we can see a direct relationship between affective distance and social media interaction. We found it surprising that comments that get more responses show more significant agreement with the content of the video, at least from the point of view of sentiment.

Discussion

In the last decades, scholars have predominantly studied affective polarization in relation to partisanship (Schieferdecker, Joly, & Faas, 2024). Particularities for the expression of citizens’ political viewpoints provided by social networks produce unique opportunities for communication and public opinion research (Hameleers, 2019). In addition, the increasing relevance and projection of new methodological approaches based on machine learning techniques (Chen, Fan, Duan, Wang, & Zhang, 2019), have grown the research possibilities in the field of public opinion research, providing a different perspective, shortening the fieldwork, and enlarging samples.

The data provided indicate the same direction as previous studies’ results: it is difficult to estimate a direct relationship between the level of network reaction and polarization, at least without moderators. In this respect, it is essential to point out that it is very suggestive that, in all the topics analyzed, the behavior of users has followed the same pattern: the comments that receive replies have a lower affective distance than the average, and their replies have a higher affective distance than the average. In other words, although there is no direct relationship between the number of comments and affective distance (and, therefore, between polarization and active participation in social media), there may be a direct relationship between responses and affective distance, albeit inversely. This result is very interesting since it is the only one observed across the sample, suggesting a pattern of behavior.

In addition, the data provided have thrown up two other unexpected points that may be linked to the above. On the one hand, the users analyzed did not behave as predicted. While other research has shown that Mediterranean users in Europe rewarded the most polarizing comments (Luengo, García-Marín, & de-Blasio, 2021), we have seen the opposite in this data. Naturally, the current sample includes Spanish-speaking Europeans, although most probably a minority (and, as we understand, without the capacity to alter the trend). Precisely, this is likely because the topics analyzed are not the same, which means that user behavior might not be independent of the topics discussed. That would mean that science, technology, and health topics may behave in different ways than political topics (Rudolph & Hetherington, 2021). However, the response comments generally tend to be more polarized, except for the science topic. That may be due to the role of clickbait in video titles (Gothankar, 2021).

On the other hand, this may also be related to the fact that including the word “science” in the title may not be attractive in itself, which explains why topics such as climate change or health generate more polarized responses, attracting users who are outside the bubble created by the algorithm. These characteristics can also affect the number of “likes” videos received and vice versa. Both positive and negative results can be related to various ideas. However, the role of YouTube’s algorithm and the criticism it has received have led to changes in the platform’s dynamics. As García-Marín & Serrano-Contreras (2023) point out, the involvement of human moderators and the gradual evolution of computational models have been a tonic to avoid extreme positions, although, the algorithm, as revealed by Ibrahim, AlDahoul, Lee, Rahwan, & Zaki, (2023), might preponderate a left-wing alignment, at least in the US.

Another unexpected finding has been that the affective distance does not increase over time, denying the sense that polarization is increasing. However, comments increase over time in almost all cases (except the science topic, which has viral videos at very precise times). Significant relationships between affective distance and time were observed in only two of the four cases; the relationship is inverse in one. That means the trend is only sometimes increasing and may even decrease. This is coherent with previous empirical evidence provided from other measures and approaches (Garzia, Ferreira da Silva, & Maye, 2023).

To conclude, we have to underline that this article does not claim to draw strong causal inferences, which is an extended limit in social science in general, and in public opinion research in particular. Some authors (Trilling, van Klingeren, & Tsfati, 2017) suggest that it is technically very difficult to model causation in this research field, with the most frequent research design proposals. However, it is quite likely that the interactions between independent and dependent variables are established in a circular dynamic, reinforcing or diminishing each other.

In any case, our data suggest that more research is needed to fully understand the role of moderators in the relationship between polarization and active participation in social media.

Funding

This research is part of the R&D&I project [PID2021-128272NB-I00 funded by the Spanish Ministry of Science and Innovation MCIN/ AEI/10.13039/501100011033/ ERDF”A way of doing Europe”]. This research has been funded by the Fundación Pública Andaluza Centro de Estudios Andaluces (ROR: https://ror.org/05v01tw04 and Crossref Funder ID 100019858), through the project PRY69/22 entitled POLAR-MED: La polarización en los medios de comunicación en Andalucía.

References

Biographical Notes

Ignacio-Jesús Serrano-Contreras is Associate Professor at the Robert Zajonc Institute for Social Studies at the University of Warsaw. He holds a PhD with international mention from the Doctoral School of Social and Legal Sciences at the University of Granada. His lines of research focus on the study of political polarization and the development of Social Sciences as an interdisciplinary axis within the construction and advancement of Natural Language Processing. Through these methodologies and fields of study, he has worked on the analysis of media (traditional and social) and political discourse in order to measure and capture social changes in issues such as feminism. He has been a visiting researcher at the universities of Hildesheim (Germany), Mostar (Bosnia), LUISS (Italy), Surrey (UK), and Bergen (Norway). He has also presented papers at both national and international conferences, as well as acting as a scientific reviewer for high-impact journals such as Comunicar and the International Journal of Sociology (RIS).

Javier García-Marín is a Professor in the Department of Political Science and Public Administration at the University of Granada. He obtained his PhD in Political Science, and a Master in Economics and International Relations. He has been a Visiting Professor at the London School of Economics and Political Science, the University of Glasgow (Glasgow Media Unit), the Amsterdam School for Communications Research, the Institut d’Études Politiques of Bordeaux and the University of Buenos Aires. He took part in many international projects.

Óscar G. Luengo is a Full Professor in the Department of Political Science at the University of Granada. He is Vice-Chair of the Research Committee on Political Communication (RC22) at the International Association of Political Science (IPSA). Director of the Escuela Iberoamericana de Altos Estudios en Gobierno Local of the Unión Iberoamericana de Municipalistas (UIM). He has been a Visiting Researcher at the European University Institute (Florence, Italy), the Johannes Gutenberg-Universität (Mainz, Germany), and the Amsterdam School of Communications Research (The Netherlands). Visiting Professor at University of California, Berkeley (USA) among other universities.

Footnotes