To preserve the integrity of their own environment, most mainstream social media platforms deploy diverse intervention strategies, targeting both inappropriate content (e.g., through flagging, demotion, or deletion) and the users (e.g., through temporary or permanent suspension) who share it. However, the effectiveness of these interventions is increasingly questioned by researchers and policymakers who demand for a proactive and holistic effort rather than the current siloed and retroactive solutions [8, 26, 34]. Indeed, even if the moderation intervention is effective on an isolated platform, it might trigger harmful activities on other platforms. For instance, Ali et al. [3] showed that, following the suspension of radical communities on Reddit, users migrated to alternative platforms becoming more active and sharing more toxic content. A similar pattern emerged following the de-platforming of Parler, where users migrated to other fringe social media platforms such as Gab and Rumble [16]. In addition, Russo et al. [29] have found that the antisocial behaviors of migrated users may spill over onto the mainstream one through other (non-radical) users active across platforms. Accordingly, Mitts et al. [22] discovered that when Facebook banned some anti-vaccine groups, the toxic content promoted by these groups resonated on Twitter. Also, during the 2020 U.S. election, videos removed from mainstream platforms were republished on the (low-moderated) Bit Chute platform [31]. The above-mentioned studies raise the need for proactive and collaborative moderation approaches to guarantee the integrity of the whole digital information ecosystem. In this paper, we investigate the potential benefits of social media platforms to share information concerning their moderation interventions by studying the users who post moderated YT videos on Twitter.

The diffusion of multimodal information across platforms (e.g. images, videos) is particularly threatening, as multimedia content has been proven to be much more attractive and credible than only textual posts [15]. In particular, the cross-posting of (harmful) video content across multiple social media platforms is a well-documented problem in the scientific literature [14, 34]. For example, a considerable number of suspicious YT videos were shared on Twitter to raise skepticism about the COVID-19 vaccination campaign [25]. Cinelli et al. [7] show that antivaccine YT videos that were shared

on Twitter experienced an increased level of visibility and dissemination on You Tube, whereas Nogara et al. [24] recognized You Tube as one of the most prolific channels used by the notorious Disinformation Dozen to spread Covid-19-related conspiracies on Twitter. Similarly, Golovchenko et al. [14] have shown that the Internet Research Agency (IRA) leveraged You Tube content in its 2016 Twitter propaganda campaign, and more recently Wilson and Starbird [34] have also reported the adoption of YT content to support anti-White Helmet operations in 2020. In general, the studies mentioned above demonstrate that harmful YT content is not only active within the source platform but can flourish in other target platforms, often with the intention of influencing vulnerable and fringe communities. In such a scenario, the entities who mobilize and disseminate harmful content can take on various forms, i.e., bots, sockpuppets, influential elites, or even information consumers susceptible to misinformation and conspiracies. In this paper, we study the characteristics of these entities focusing on Twitter users who share moderated YT videos and investigating their similarities and differences with the users who share nonmoderated YT content.

We used a dataset of election-related tweets collected using Twitter’s streaming API service in the run-up to the 2020 US election [5]. In particular, we focus on the six months, from June 2020 to December 2020, covering the last part of the electoral campaign, as well as the aftermath of the election, which was characterized by the widespread diffusion of misleading claims and conspiracies on the integrity of the election results [12, 32]. In this period of observation, we collected more than 600M tweets (including original tweets, replies, retweets, and quotes) shared by 7.5M unique users.

In particular, tweets that include YT videos account for 0.65% (3.9M) of the collected messages. Note that we do not consider URLs to YT channels. Fig. 1 shows that the fractions of tweets (resp. users) sharing YT content are consistently larger than tweets (resp. users) pointing to other mainstream social media, which is in line with [1, 6]. In addition, in both cases, we observe an increasing trend towards our observation period (the second half of 2020), possibly because of the approaching election (November 3rd, 2020). In total, 527k YT videos were shared on Twitter by 830k users. Through the You Tube API, we could retrieve various video metadata, including the ability to determine if a particular video was removed from the platform. We find that, among all YT videos, 24.7% (130k out of 527k) were moderated. Furthermore, before the intervention, these videos were shared on Twitter by 34.5% (287k out of 830k) users. Finally, we collect YT video metadata of nonmoderated videos, including the video title, description, tags, and the channel that published the video. Interestingly, You Tube does not allow one to collect metadata for moderated videos, including the date when the intervention occurred as well as the reason(s) for the moderation intervention. However, we can safely assume that a video is still online when shared in an original tweet, as this requires the user to report the video URL in the Twitter post.

To identify the mobilizers of moderated YT videos, we first consider the most active YT mobilizers, as our objective is to investigate the characteristics and behaviors of users who repeatedly (rather than occasionally) post YT content. For this reason, we consider users who shared at least 5 YT videos on Twitter, which results in a set of 113k Twitter users. Then we partition this list of YT mobilizers into two groups based on the volume of shared moderated YT videos. For each user u, we define the ratio of moderated videos (rmv(u)) as the proportion of moderated YT videos out of the total number of YT videos shared by u during our observation period. Based on this metric, we define nonmoderated YT video mobilizers (NMYT)

Figure 1: The monthly percentage of tweets (left) and users (right) who shared a link to a mainstream social media platforms in 2020

5000

0

0.0 0.2 0.4 0.6 0.8 1.0

rmvratio

(a)

To answer RQ1, we perform an analysis of the consumption of YT videos on Twitter comparing moderated vs. nonmoderated YT videos. It is worth noting that moderated videos tend to be shared for

% of suspended users

60

50

40

30

0.0 0.2 0.4 0.6 0.8 1.0

rmvratio

(b)

a limited number of days (20 days on average), while nonmoderated videos have a longer lifespan (50 days on average), likely due to YT moderation interventions. Therefore, to perform a fair comparison, we examine the number of tweets including the YT videos during the first week after its first share on Twitter. Fig. 3a shows the distributions of the number of tweets for nonmoderated and moderated videos. It can be noted how moderated videos distribution is characterized by a right heavy-tail, meaning that moderated videos when posted for the first time on Twitter, generate a higher volume of sharing activity with respect to nonmoderated videos.∗This finding is consistent with the previous study by Locatelli et al. [19], who analyzes Covid-19 YT videos and shows that moderated videos prompt more active engagement from viewers. To further explore the prevalence of moderated YT content on Twitter, we compare the volume of interactions with content originating from other social media platforms. Specifically, Fig. 3b and Fig. 3c show the volume of tweets and retweets of moderated YT videos, URLs pointing to mainstream online social networks, and URLs redirecting to fringe platforms [33]. We observe that the volume of tweets linking moderated YT videos alone is greater than

∗A Mann–Whitney test (p-value< 0.01) was performed to validate this finding.

Figure 2: YT Mobilizers characteristics with (a) Distribution of the rmv(u); (b) Percentage of the suspended users with respect to their rmv(u). The shaded area is the 95% confidence interval.

(a) (b) (c)

Figure 3: The prevalence of moderated YT videos with: (a) The distribution of the number of tweets sharing each video during the week after its first share; (b) Number of original tweets containing a link to each social media platform (Log-scale); (c) Number of retweets containing a link to each social media platform (Log-scale)

original

quotes

1.00

1.00

0.75

0.75

0.50

0.50

0.25

0.25

0.00

0.00

NMYT MYT

NMYT MYT

user group

user group

the volume of tweets pointing to any other social media platform. According to [6], we find that fringe content supplied by Parler and Bit Chute is outnumbered by the content provided by mainstream platforms.

Findings and Remarks. Addressing RQ1, we discovered that moderated You Tube videos were widely shared on Twitter during the run-up and aftermath of the 2020 US election. In addition, moderated videos received a higher volume of interactions in the early days of their lifespan. Additionally, moderated YT content alone received more engagement than whole content from other mainstream platforms.

To address RQ2, we characterize YT mobilizers that share moderated (MYT) and non-moderated videos (NMYT) and investigate whether these users show significantly different behaviors and characteristics across three dimensions:

• Cross-Posting Activity: we explore the sharing activity that users perform on Twitter, including their cross-posting of content originating from other mainstream and fringe platforms;

• Trustworthiness of the account: we examine whether MYT and NMYT users are verified accounts or bots, also looking at their account status (active vs. suspended) and potential involvement in information operations;

• User Interests: we investigate the political leaning and topics of interest of mobilizers, both on Twitter and YT.

4.2.1 Cross-Posting. We consider all the possible sharing activities that a user can perform on Twitter, i.e., posting an original tweet, commenting on a tweet with a reply, re-sharing a tweet with or without a comment via retweet or a quote, respectively. From Fig. 4 it can be seen that while exhibiting similar behaviors in posting original tweets, MYT users retweet more and reply less than NMYT users. To further characterize this discrepancy in the use of retweets, we examine the proportion of re-shared tweets embedding links to other web domains (beyond You Tube). We find that MYT users often retweet content linking to external sources, i.e., 50% of their retweets on average contain URLs with respect to 28% for NMYT

reply

retweet

1.00

1.00

0.75

0.75

0.50

0.50

0.25

0.25

0.00

0.00

NMYT MYT

NMYT MYT

user group

user group

mobilizers. This analysis suggests that the activity of NMYT users is more diversified on Twitter, while MYT users tend to passively consume and spread (through retweets) the content they see on Twitter, especially if such content links to external resources. To further characterize the two groups of mobilizers, we consider how they engage with YT videos on Twitter. Specifically, we aim at understanding whether users passively interact with YT videos (e.g., through retweets) or proactively share them (e.g., in their original tweets). To this end, we define users as producers (resp. consumers) of a YT video if their first share of that video is an original tweet or a reply (resp. retweet). Then, we define prodratio as the proportion of YT videos produced by a user out of the total number of videos with which he/she engages. It is worth noting that with “video producer” we do not imply that the user is the publisher of the video on YT. Interestingly, we find that the mobilizers in both groups are either mostly producers (prodratio > 80%), or mostly consumers (prodratio < 20%) of YT videos. Specifically, we count 15,761 (62.1%) producers and 4,439 (17.4%) consumers in the NMYT group, while there are 4,630 (31.9%) producers and 3,742 (25.8%) consumers belonging to the MYT group. Fig. 5a shows the distribution of the prodratio in both groups. On the one hand, we find that NMYT mobilizers are mainly producers, that is, the 25 percentile of the prodratio distribution is 0.50. However, the MYT group includes the same number of producers and consumer accounts. This result is consistent with our previous finding and confirms the tendency of MYT users to passively retweet content compared to NMYT accounts, which actively participate in the discussion through replies or quotes. Furthermore, we investigate how mobilizers engage with content from other social media platforms. In particular, we target mainstream platforms, i.e., Facebook, Instagram and Reddit, and the fringe platforms outlined in Fig. 3b. We define the extreme ratio as the fraction of extreme URLs out of the total number of tweets shared by each user. Fig. 5c shows that the MYT and NMYT mobilizers have similar distributions in terms of mainstream ratio (p-value > 0.01 with a Mann-Whitney test). Taken together, the MYT and NMYT mobilizers share the same percentage (3.3%) of mainstream content on average, and no user posts more than 15% of tweets linking other

Figure 4: The distribution of original tweets, replies, retweets and quotes for NMYT and MYT mobilizers

mainstream social media platforms. However, Figure 5b shows that MYT users engage with fringe platforms much more than NMYT mobilizers. In fact, the distribution of their extreme ratio is different according to a Mann-Whitney test (p-value< 0.01), even if they have similar mean values, that is, 0.5% and 0.7% for NMYT and MYT mobilizers, respectively. In general, this analysis highlights that the two groups of mobilizers do not show any difference when interacting with mainstream social media platforms, while MYT mobilizers share more content from low-moderated online spaces than NMYT users, suggesting a form of endorsement of extreme ideas pushed on fringe platforms [33].

4.2.2 Trustworthiness. Here, we investigate the nature and status of the accounts in the two groups of mobilizers. Given the pivotal role that political elites, bot accounts, state-backed trolls have in orchestrated campaigns and (mis)information operations [10, 14, 21, 24, 35], we aim at identifying the entities pushing moderated or nonmoderated YT videos on Twitter. To this end, we use Botometer [36] and Twitter API to assess whether our mobilizers are automated or verified accounts, respectively. As shown in Figure 6, we find that both groups include very few verified accounts (268 and 19 in the

NMYT and MYT groups, respectively) and bots (2, 234 and 586 in the NMYT and MYT groups, respectively). Next, we examine whether users in the two groups of mobilizers were suspended by means of Twitter moderation interventions. Indeed, during the 2020 US election, the platform made an increased effort to ensure the integrity of discussion by adding warnings to suspicious or misleading content, as well as suspending accounts involved in information operations [20, 30]. From our perspective, we are interested in understanding whether and to what extent the accounts of the MYT and NMYT mobilizers were moderated by these actions. We find that accounts are suspended in both groups but in different proportions. Fig. 6 shows that 53.8% of MYT mobilizers (7,793 accounts) are moderated by Twitter, while 31.4% of NMYT users (7,984 accounts) are suspended. In addition, we further explore whether the (suspended) accounts are involved in state-backed information operations (in short Info Ops) on Twitter. As reported in Figure 6, we find that a minority of mobilizers (599 accounts in total) are involved in these campaigns and, interestingly, NMYT users are more involved than MYT users (2.2% of MYT against just 0.2% of NMYT). Overall, this analysis suggests that

(a) (b) (c)

Figure 5: The distribution of prodratio, extreme ratio and mainstream ratio for NMYT and MYT mobilizers

Figure 6: The number of accounts in each mobilizer group that were verified, bots or suspended. The columns are as follows: “Total Accounts” is the total number of accounts in each group. “Total Videos” is the number of unique YT videos shared by each group. “Verified Accounts” is the number of verified accounts in each group. “Bot Accounts” is the number of accounts labeled as a bot by the Botometer API in each group. “Suspended Accounts” is the number of accounts in each group that were later suspended by Twitter. “Info Ops Accounts” is the number of (suspended) accounts involved in information operation in each group

Table 1: Most shared hashtags and YT video keywords by NMYT and MYT mobilizers

NMYT Mobilizers MYT Mobilizers

#putinspuppet #krakenteam #trumpvirus #chinebitchbiden #resignnowtrump #demonrats #trumplies #evidenceoffraud #traitorinchief #bestpresidentever45 #gojoe #arrestfauci #trumpkillus #trumpwon #weirdotrump #trumppatriots

Twitter hashtags

barkhuff dan rsbn bernie sander bobulinski lincoln censored rainbow christina bobb loyalty fitton cnn spoiled incompetence tucker carlson

You Tube keywords

even though MYT mobilizers violated Twitter policies, they are not involved in state-backed orchestrated campaigns during the election.

4.2.3 User Interests. We now turn our attention to the content shared by MYT and NMYT mobilizers. As the data under analysis relate to the U.S. 2020 Presidential election, we expect that the discussion resonates with political topics, especially regarding the electoral campaign of the candidates, as well as the heated discussion around the alleged evidence of fraud during the aftermath of the election. Also, we focus on the political orientation of the users under analysis and on their potential connection with conspiratorial and fringe theories. To explore the general interests of MYT and NMYT mobilizers, we compare the hashtags of their tweets, as well as the descriptions of the YT videos they share on Twitter. To this end, we apply SAGE [9] to find the most distinctive hashtags and keywords for the tweets and video descriptions, respectively, shared by the two groups. It is worth to note that for MYT users we only took into account the non-moderated videos as the You Tube API cannot retrieve metadata related to moderated content. Table 1 reports the keywords and hashtags extracted by SAGE. We observe that MYT users are Trump supporters, as shown by the hashtags #bestpresidentever45, #demonrats, and they also sustain his allegation of voter fraud after the electoral count, as seen by the hashtags #krakenteam, #trumpwon. On the contrary, NMYT mobilizers explicitly despise Trump, as can be seen by the hashtags #trumpvirus, #traitorinchief, but their political orientation is not as clear as for MYT users. Indeed, #gojoe is the only pro-Biden hashtag in the NMYT’s top-50 hashtags. Overall, the keywords extracted from the descriptions of the YT videos align with the Twitter hashtags of the two groups. However, they do not communicate positive or negative sentiment but usually refer to (groups of) people who publicly stated their political preferences. For example, NMYT Mobilizers shared several videos related to the lincoln project and its ad starring Barkhuff Dan explicitly

saying “I can see Trump for what he is — a coward. We need to send this draft-dodger back to his golf courses”. It is worth noting that the Lincoln project is run by Republicans opposing Trump, which supports our intuition that NMYT users are not always Biden supporters. On the contrary, the MYT mobilizers support for Trump is clear also from their shared videos mentioning Christina Bobb, who has been close to Trump’s legal team that tried to overturn the result of the Presidential election, and Tucker Carlson, who has recently been nicknamed as Trump’s heir.†

To further investigate the political orientation of users, we leverage the political leaning score assigned by Media Bias Fact Check to several news outlets and, consistent with previous work [12, 27], measure user political orientation by averaging the scores of the domains they share on Twitter during the observation period. Fig. 7a and 7b show the top-10 domains shared by YT mobilizers and the political leaning distribution of MYT and NMYT users. The former group includes several far-right users (the median of the distribution is 0.47) who mostly share news from breitbart.com and thegatewaypundit.com, which are known to promote conspiracy theories and publish extreme conservative content [18, 23]. In contrast, the latter group includes less extreme and more liberal users. However, the political lean distribution of the NMYT group is bimodal (the larger mode is −0.27 and the smaller one is 0.41) and a small subset of these mobilizers shows an extreme conservative ideology. This result is further confirmed by looking at the top-10 domains of MYT mobilizers, including Foxnews.com, which is also shared frequently by the NMYT group, and forbes.com, which is a center-right news outlet.

Findings and Remarks. In response to RQ2, we found that MYT users tend to passively retweet what they see on Twitter rather than actively posting original tweets or replies. In addition, they are usually suspended on Twitter but are not involved in information operations. Finally, when assessing the (political) interests, we found that MYT are far-right supporters and backed Trump during the 2020 US election, while the political leaning of NMYT users is less extreme and more diverse.

In this article, we studied the Twitter discussion around (video) content that is deemed harmful on You Tube. Leveraging an unprecedented large-scale dataset of 600M tweets shared by more than 7.5M users, we discovered an unexpectedly high number of

Figure 7: (a) The news outlet shared by each group of mobilizers (we omit the .com extension for brevity) ; (b) the distribution of the political leaning within the two groups of mobilizers

Figure 8: Interaction patterns enacted by NMYT and MYT accounts: (a) Proportion of interactions between You Tube mobilizers normalized by the source; (b) Proportion of interactions between You Tube mobilizers normalized by the destination; (c) Z-scores of observed retweets between You Tube mobilizers (p-value < 0.01)

moderated YT videos shared on Twitter during the 2020 US election. Overall, moderated videos were shared more than nonmoderated ones and received far more attention than content from fringe social media platforms. Moving beyond previous work, we investigated the characteristics of the Twitter users responsible for sharing both moderated and non-moderated YT videos. On the one hand, we found that users sharing moderated content tend to passively retweet what they see on Twitter rather than actively posting original tweets or replies. On the other hand, the mobilizers of nonmoderated videos actively share YT videos in their original tweets. Overall, most of the users were regular Twitter accounts rather than bots or state-sponsored actors, and, even if we did not find any involvement in information operations, Twitter suspended more than half of the moderated video mobilizers. Furthermore, we found that the mobilizers of moderated YT videos are far-right supporters and sustained Trump during the 2020 US election. By contrast, the political preference of the mobilizers of non-moderated YT videos is more diverse since users in this group range from Biden supporters to other Republican representatives who did not endorse Trump’s political campaign. Finally, we studied the interactions between the mobilizers of moderated and non-moderated videos and discovered that both groups exhibit strong group cohesion and are engaged similarly to the general Twitter audience.

There are a number of limitations to our study. First, neither Twitter nor You Tube provides any additional information on account suspension and video moderation, and the timing of their interventions is also unknown. Therefore, there is no guarantee that YT videos were still online when reshared through retweets on Twitter, but we can confidently assume they were not moderated yet when shared in an original tweet. Furthermore, we acknowledge that our analyzes, as in several previous works [17, 19], could be biased towards moderated YT content that includes not only videos that violate You Tube policies but also those removed by their publishers for any reason. Second, we overlooked the YT channels shared on Twitter to safeguard our analysis from Twitter users who just advertise their own (or others) You Tube channel [2]. However, this choice might prevent us from considering another potential source of harmful YT content on Twitter. Third, the partition strategy to define the two groups of mobilizers is quite conservative, since we considered Twitter users who never share moderated videos and those who mostly share moderated videos.

Our study has two major takeaways: first, moderated YT videos are widely shared on Twitter, and users who (passively) share those endorse extreme and conspiratorial ideas. From a broader perspective, we have shown how harmful content originating in a source platform significantly pollutes discussion on a target platform. Although more research is still needed, we conjecture that sharing information about the interventions taken would improve our understanding of cross-platform harmful content diffusion and benefit all entities within the information ecosystem. For instance, in the You Tube-Twitter cross-posting scenario considered in this paper,

You Tube moderation activity can benefit both parties of the cooperation: on the one hand, Twitter has the opportunity to (early-)detect intra-platform harmful activities; on the other hand, You Tube can further improve its moderation based on the cross-platform signals tied with harmful YT content diffusion on Twitter. Second, the mobilizers of the moderated YT videos appeared to be regular Twitter users who do not necessarily share content from fringe platforms. This suggests that cross-posting (harmful) cross-platform content is participatory [20] and research in this field should not only target bots and trolls but instead consider the role of online crowds and more complex social structures on different social media platforms. Future work might build upon our findings to design algorithms to automatically identify or predict whether a YT video will be moderated based on the engagement it receives on Twitter, as well as to detect early signals of radicalization. In addition, we aim to investigate whether our results generalize to other topics beyond political elections or other highly-moderated social media (e.g. Facebook, Instagram).