Malicious Earworms: How the far right surfs on TikTok audio trends

Team Members

Facilitators and designers: Marloes Geboers and Benedetta Riccio

Canjie Kuang, Eileen Yang, Hille Steenhuis, Jiaye Liu, Kexin Xiong, Pawel Bitel, Macey Mino, Ruijia Song, Nicole Nobel, Yaxuan Wu,Riccardo Freschi

POSTER 1

POSTER 2

Contents

1. Introduction

In December 2024, the Dutch NCTV (Nationaal Coördinator Terrorismebestrijding en Veiligheid) published its yearly ‘Dreigingsbeeld Terrorisme Nederland’ which establishes current threat levels and where it was communicated how a substantial level is maintained, as some hundreds of people (from age 14 to up) have been seen to be active in online rightwing extremist milieus where they not only share their sympathy for hateful ideologies, but also scheme violent activities and recruit others. In Germany, parallel developments unfolded over the last (couple of) years. Fuelled by the 2024 state elections in Thüringen and in Sachsen-Anhalt and a fatal knife attack in Solingen in August, far right extremist ideas embedded or otherwise linked to the political campaign of the Alternative für Deutschland (AfD) party, have been amplified by accounts on TikTok that participate in far right propagandist messages. In a report by the ISD (Institute for Strategic Dialogue) it was found how Neo-Nazis and white supremacists are sharing Hitler-related propaganda and trying to recruit new members on TikTok.

The TikTok algorithm is (inadvertently) promoting this content to new users, as extremist communities are leveraging the popularity of TikTok among younger audiences to spread their message. This project aims to address how such accounts operate, more specifically we address how extremist TikTokers navigate the specific affordances of the app’s sound architecture and its subsequent user pathways that, as this research will lay bare, hosts significant loopholes that can be exploited by malign actors, either surfing on benign audio trends or linking racist songs to popular meme templates.

Sound and speech moderation

While rapid advancements in LLM’s capacities to detect hate speech start to encompass context-dependent covert language (Hee et al, 2024; Steen et al., 2023), sound affords extremely ‘subtle’ tactics of circumvention that are ubiquitously present on TikTok. Moreover, speech is demanding to moderate as it “has to be collected, transcribed, parsed and interpreted; all of which demands additional effort and expertise” (Kevin Jin, 2024). It is clear that audio presents a fundamentally different set of challenges for moderation than text-based communication: A report from ICUC outlines how “Audio-based social channels–TikTok, Instagram’s Reels, Clubhouse, Discord’s Stage Channels–started to research solutions, but the tools for audio content moderation are lagging far behind social listening tools for text-based conversations. A few companies are developing speech analysis API, but there is no streamlined approach.”

While TikTok prioritizes physical and mental well-being and above all physical safety in their community guidelines (Christin et al., 2024), they also state how ‘promoting hateful organizations and hateful ideologies’ is: ‘NOT ALLOWED’ (Combating hate and violent extremism). From 2020 onward the platform has further specified categories on which they (say they) will take action. The archived community guidelines reveal how in 2019 the platform outlined a generic description of not promoting violence targeted at people based on protected attributes. Such attributes have, overtime, come to be specified into racial supremacy, misogyny, anti-LGBTQ+, antisemitism, and most recently added Islamophobia.

Especially pertinent in light of the data that we will be working with, is the explicit statement that TikTok prohibits: “dehumanizing someone on the basis of their protected attributes, such as calling them criminals, animals, or comparing them to inanimate objects”. Ongoing research into far right spaces across the UK, the Netherlands and with a focus on Germany, point to a presence of posts doing just that, and while these posts usually get indexed within dispersed, low-engaged-with ‘original sounds’, they do find a home in these sound networks that connect members of a ‘minority’ that is nonetheless growing and radicalizing.

2. Research Questions

The overarching question was: How is TikTok ’s communicative sound architecture affording the proliferation and obfuscation of problematic content?

Sub questions (sub projects):

#Project on Soundscapes of German and Russian Marches (poster #2).

Departing from examining both avatars (profile images) as well as the content they post in connection to the soundscape used as its background, we asked ourselves several questions:
Can a certain TikTok sound be connected to users’ extremism alignment?
If so, could an overlap exist between the fields of interest of certain groups of people and being more prone to post extremist content?
If a certain song used as a background for a post contains potentially extremist lyricism, will that find its reflection in the extremist nature of a post?
Additionally, since we’re looking at both the German and Russian side of TikTok in connection to rather particular periods, what will be the prevailing type of content, and how much of it could be considered extremist?

#Project on the Story of the Motorfans (poster #2).

What is the role of distributed 'original sounds' in perpetuating far-right, white supremacist,
and xenophobic content on TikTok?
In this context, we propose the following hypotheses:
1. TikTok 's algorithmic system tends to promote frequently reused sounds, inadvertently
amplifying the visibility of extremist content.
2. Regular users, when replicating or participating in challenges created by extremist
accounts, may unknowingly propagate "original sounds" embedded with extremist
messaging.

#Project on visual analysis of Love Nature, Hate Antifa (poster #1).

What role do soundscapes and templatability play in the modulation of participatory propaganda on TikTok?

#Project on Soundbangers and Nationalist Sentiment (poster #1).

How does TikTok 's communicative sound architecture facilitate the reverberation of nationalist content within user comments?

3. Methodology and initial datasets

The presented study takes as its focus the networked linkages between the AfD campaign and extremist rightwing niches on TikTok that are ‘held together’ and network through sounds. These sounds can be hijacks of popular songs that through association start to map onto a political ideology, but they can also be blatantly problematic in their contents. In the latter case, such sounds tend to latch onto meme templates in order to attain visibility.

#Project on Soundscapes of German and Russian Marches (poster #2).

Datasets we worked with:

Sound pages:
“Farewell of Slavianka” [“Прощание славянки”]
“Erika”
“The Devil’s Song” [“Teufelslied”]
“Where all roads end” [”Wo Alle Straßen Ende”]
“The Sacred War” (“Священная война”)

The process of coding was divided into two parts. First, to outline the poster's potential online environment or field of interest, the avatars of the TikTok users who posted using the soundscape have been examined. At this stage, the avatars have been coded for the potentiality of extremism (Y/N). This categorization mostly involved looking out for national symbols concerning militarisation, overtly nationalistic, or offensive symbolism. Subsequently, the content itself was also taken into consideration. This involved reviewing the collected posts and coding them into categories according to their content. The categories were chosen after briefly revising the posts to find potential groupings and fields of interest that are more or less shared amongst multiple posts. As per the explanation of the rationale, we regard as memes everything that is either made with a humoristic intent or follows an established meme trend. "Historical" is a category for educational or historical posts that are relatively objective and discuss past events or objects. "Geopolitical" is for all posts that are either invested in current military conflicts, are political, or are opinionated on different political situations, groups, or nations. "Games" involves mostly video game content and other games, such as modeling or reconstructions. "Face" is content that could also be connected to ideas delineated in the "Geopolitical" category but emphasized through the intentional connection of the topic to the user's

identity through showing their face. This involves all sorts of opinion pieces or short vlogs. "Other" involves content that, despite using the song as background, could not be classified under other categories. Finally, additional categorization has been given to the posts that are overtly or potentially extremist (with the same rationale as the users' avatars), and they have been given an "E" category (for "extremism") next to their already assigned one.
Finally, to visually distinguish certain trends, including possible shared fields of interests of the users identity as well as the overarching trends of their posts, the images for the avatars have been fed into ImageSorter (visual-computing.com, 2024), which generated images showing visual clusters of both datasets. For that purpose, the German and Russian songs were merged and portrayed individually. Conversely, the results of the coding of the posts have been portrayed in the form of pie charts for readability. This includes the total number of posts on both sides (Russian and German).

#Project on the Story of the Motorfans (poster #2).

This study employed digital methods, focusing on the popular music named “Anotha European classic” and its associated video content on the TikTok platform, including analyzing the co-tag network and account network. In order to collect the data, we used the software Zeeschuimer to scrape content from TikTok and
uploaded it to the software 4CAT for analysis (Peeters and Hagen 2022). In the data collection phase of our study, we employed a two-stage sampling strategy: Firstly,
we searched for the original sound “Anotha European Classic”, which is based on a techno song named Ferrari. Subsequently, we utilized the Zeeschuimer tool to scrape data within this soundscape on TikTok, gathering a total of 1,154 video entries. To ensure the feasibility of our data analysis, we then applied the random sampling method, uploading the data set to the software 4CAT to select a random sample of 400 videos for further analysis.

Data analysis

To investigate how accounts with varying political stances distribute and interact around this
specific audio content, we employed a systematic analysis using Grounded Theory coding
methods on the collected data. The coding process was divided into three stages: open
coding, axial coding, and integrating categories. Specifically, in the open coding phase, we
thoroughly watched and analyzed each video sample to label, identify, and construct
preliminary categories. During the axial coding stage, we reviewed video data and extracted
more appropriate categories and constructed associations and logical links among them.
Throughout this process, we paid particular attention to elements and themes in video content
that might be associated with far-right ideology. Finally, by integrating the identified
categories, we gained deeper insights into the specific role of "original sound" in the
dissemination of far-right content.

To better understand the relationships between categories, we utilized Gephi software to
construct a co-tag network. This network visualization approach enabled us to identify
content clusters and examine how different clusters interconnect within the context of this
particular sound's usage. Additionally, we conducted in-depth ideological coding of accounts
within specific categories, categorizing them based on their political orientation (e.g.,
left-wing, far-right), and created an account network to visualize these ideological
connections. This dual-network analysis approach provided comprehensive insights into how
accounts with different ideological stances interact around this specific sound and how
extremist content potentially circulates between different ideological communities. By
combining content analysis with network analysis, we were able to reveal the mechanism
through which this sound functions as a medium for far-right content dissemination on
TikTok.

While our random sampling of 400 videos from a total of 1,154 entries ensures analytical
feasibility, it may limit our ability to capture all significant variations in content and usage
patterns. Furthermore, our ideological coding method presents certain limitations. The
identification and categorization of political stances can be inherently subjective, particularly
when analyzing content that deliberately obscures its ideological positioning. The
interpretation of multilayered meanings in video content presents considerable challenges,
especially those employing subtle cultural references.

#Project on visual analysis of Love Nature, Hate Antifa (poster #1).

We started by investigating different soundscapes on TikTok to find out whether certain patterns or templates could be detected for a specific sound. The sound we decided to research is named ‘оригинальный звук’, which translates to ‘I am your nightmare’ and is an original sound version of the Russian song ‘Австралия (Australia)’. The videos under this sound showed a significant amount of antifascist and fascist content, so we then scraped them using Zeeschuimer. Since there were a lot of videos in the dataset, we then used 4CAT to generate a randomly compiled dataset consisting of 400 videos. These were then coded and placed into different categories based on the content and trends we could detect. The categories are: LNHA (which stands for ‘love nature hate antifa’), LNHA Variations, LNHA Counter, Political, and Song Lyrics. After the coding process, the comments of the top commented video for each category were scraped. We then uploaded the comments to labs.polsys.net to create bigrams, which were then converted to a CSV file. All comments that appeared more than once were filtered, narrowed down, and finally uploaded to RAWGraphs 2.0 to create a Matrix Plot (Figure 1). We furthermore created frame-by-frame timelines for the top 5 commented videos from each category using 4CAT

#Project on Soundbangers and Nationalist Sentiment (poster #1).

Our research focuses on the soundscape of a widely known club song L’Amour Toujours by Gig d’Agostino in 1991, investigating how nationalist sentiments are evoked and expressed through comments from a highly engaged post within this soundscape.
4
Data collection and coding
First of all, we utilized the digital research tool Zeeschuimer to extract video posts within the chosen soundscape. The initial dataset contains 603 TikTok video posts published from October 2023 to January 2025. Secondly, due to the large number of diverse topics in these posts, we adopted a systematic coding process on the initial data through close reading of each post and carefully analyzing the content to identify recurring themes or implicit messages. Based on these observations, we categorized the posts into three distinct groups, which are political, religious, and entertainment. Within the political video posts, contents are predominantly focused on issues such as immigration in Europe, right-wing movements, and extremist ideologies related to Nazism. Since the majority of the video posts are politically oriented, and nationalism-related content is closely intertwined with political themes, we further narrowed our research focus to posts specifically related to political topics. Thirdly, we input these political posts into 4CAT and counted out the top 100 posts ranked by the number of comments, which indicates active users’ engagement within these posts. Finally, the post with the highest number of comments was selected as the subject of this research, which contains 12.6k comments. This post is presented in the form of two slides. The first slide asks viewers about their nationality, while the second slide showcases a ranking system created by the blogger.

Analysis method
According to Alafwan (2023), analyzing comments on social media platforms offers a valuable data source for investigating implicit insights about users, posts, categories, and community interests, while it also facilitates understanding the deliberative potential of emerging technologies and their ability to foster a virtual public sphere. Due to the larger amount of comments from the chosen post, we conducted three processes for the classification of comment content, content analysis of comments, and visualization of the final outcome. By utilizing the digital research tool 4CAT again, we identified the six most frequently mentioned countries in the comments, which are Germany, Poland, Romania, Italy, Dutch, and Hungary, and categorized them into six main groups. Subsequently, we manually assigned the remaining comments to the respective country categories based on their contents. Since these comments not only contain textual messages but also include emojis, we utilized the content analysis method to categorize them, which is a technique used to identify various types of data, including visual and verbal information, and can facilitate the organization of phenomena or events into specific categories, enabling more effective analysis and interpretation (Harwood & Garry, 2003). Finally, we input our data into Raw Graph, which is a website that can produce multiple charts. Through the Circle Packing function, our data is visualized in a hierarchical structure. This visualization represents the frequency of mentions for the six countries through the size of the outer circles. Within each country's category, the size of the inner circles reflects the frequency of related comments.

4. Findings

#Project on Soundscapes of German and Russian Marches (poster #2).

In order to visualize members of different online communities post content under the same soundscape, we
decided to merge the collection of TikTok profile avatars and present them in the form of a image sorting clusters.
The results were scraped from posts using two prevalent war marches sung during the Soviet era as soundscapes –
“Farewell to Slavianka” and “Sacred War”. The results show a number of distinct clusters of people coming from
different communities or presenting themselves differently through TikTok avatars. There is a large quantity of user
avatars consisting of national symbols such as flags from both the current as well as the Soviet era of Russia. This is
visible in both the blue and red part of the visualization starting in the middle and moving towards the right. There is
a significant presence of people having avatars that could be described as historical as seen in the top-left and
bottom-right parts of the visualization, consisting of portraits of numerous historical characters. Other than
numerous miscellaneous groups, such as stylized or animated pictures (partially visible in the bottom left corner),
there is also a number of people using their own face within this space, both masked (bottom left part), as well as
unmasked (top part moving towards the right). When it comes to images that may skew towards extremist
representations, these are present mostly within the national symbol section, with numerous hammer and sickles,
“V” symbols. The controversial “Z” symbol also appears there, but due to its changing coloristic it is also seen in
both the top-left and right section.

Crop from Poster #2 (see link on top of page).

The Alluvial diagram highlights a key trend we found while browsing TikTok through historical German march
landscapes. We found that although most of the users did not have a profile picture containing potentially extremist
symbolism, through the usage of a nazi march song called “Erika”, a memetic format was born on the platform. This
meme format consists of users posting content containing nazistic symbolism such as swastikas, Hitler lookalikes
and subjects heiling. Some of the videos portray a sort of accidentality, users “stumbling upon” nazi symbols in the
wild - marked in the diagram with a “No” in the third step; some of the content is portrayed and arranged more
deliberately - marked in the diagram with a “Yes” in the third step. Nonetheless we conclude that the usage of such
blatant nazistic song for memetic purposes poses a risk for the platform and its users, where soundscapes lead to
rabbit holes due to the high overlap of the different types of content found on TikTok, as seen in the second step of
the Alluvial diagram. Curious users beginning their journey on a meme or a war game video are only a couple of
clicks away from discovering more extremist content, such as political testimonials.

#Project on the Story of the Motorfans (poster #2).

Our research focused on the soundscape of "Anotha European classic," investigating the role of distributed 'original sounds' in perpetuating far-right, white supremacist, and xenophobic content on TikTok. While named Anotha European Classic on the platform the sound is based on a techno song named Ferrari. It assembled 4718 posts that used this ‘original sound’. The sound in isolation is not far right, but many problematic accounts and contents are linked to it. To see how the political stance of accounts distribute
across topics that link to the sound, we adopted a coding strategy, comprising open and axial coding. This resulted in thirteen categories: Political, Military, Gender
Issues, Social Issues, Racial, Religious, Cultural, Subculture, Entertainment, History, Immigrant, Economic, and Motorcycle. To explore the relationships between categories, we constructed a co-tag network. This reveals the presence of motorcycle fans who mostly, after qualitatively checking out the accounts, seem to be benign motor fans using the sound.

The fact that these motor fans use a sound that is used by many far right accounts might not be fully coincidental. The love for motorbikes was tapped into by an AfD promotion poster where party leader Björn Höcke is sitting ‘very demure’ on a signature brand motorbike that evokes many TikTokers with nostalgia or love for the country to post their motorbikes on the platform. This visual meme then pulls in both far right accounts (red nodes in network) as well as many more benign motorists (green in network) who might not have a clue that they are sharing the soundscape with far right people. This unconscious connection in the meantime expands the affective reach far right sentiment beyond explicitly political spaces. When a 13-year old with a cool bike picture slideshow sees other bike posts and decides to use the same sound, this person’s posts gets networked through the sound with far right users.

Snippet of Poster #2, link on top of page.

#Project on visual analysis of Love Nature, Hate Antifa (poster #1).

Many posts using ‘Australia’, a 1997 pop song, used a meme that was networked through the sticker text stating
‘Love Nature, Hate Antifa’. This templated ‘masterplot’ spawned many variations that were not all politically infused.
We kept the variations that either reverberate the fascist message and the variations that counter fascists and
mapped the thumbnails of the videos in an image wall where the overlaid images reflect the counter fascists who
are stating for example love nature, hate fascism. Outside of the meme template, there were posts with a clear
political message and posts that literally reflect the song in the visual layer of the posts. We were interested in how
many posts were actively using the meme template to counter the far right message and in a second analysis
(matrixplot below) we show how common comments bigrams (co-occurring words, x axis) distribute across the
political stances within the templated posts and across political and song-related posts (y axis).

The soundscape captures posts that tap into the Love Nature Hate Antifa meme template. This invites imitations
(variation below) as well as counterimitations (counter variations). We selected top commented posts for the three
categories, a political post and a post about the sound. The post boasting the initial meme displays the most
overlap in comments across categories. The political post was a slideshow featuring countries, reverberating in
comments signaling nationalities. The most frequent comments related to the song lyric post were repeating lyrics
shown in the video relating to people wanting to leave the country.

Snippet from poster #1, link on top of page.

#Project on Soundbangers and Nationalist Sentiment (poster #1).

This study primarily employs comment analysis on the selected post to uncover the patterns of interaction, underlying tendencies, and potential cultural implications within this specific soundscape.This post serves as an interactive engagement platform where users participate by commenting on their respective countries, while the author assigns ratings based on their subjective perceptions of these nations. Analysis of the comments reveals that the majority of engagement originates from six key countries: Poland, Germany, Romania, Italy, the Netherlands, and Hungary.

Comment Analysis
Some friendly emojis and the author's expressions of goodwill toward each country in the graph can be observed. In the comments, we observed that while the author frequently expressed positive sentiments such as "I love xxx (country)" or "I really like xxx (country)," these remarks were often accompanied by a subjective ranking system for various countries. For example, Belgium was rated as 🟦, Poland as 🔲, and Norway as 🟪.This pattern suggests that the author’s engagement went beyond simple expressions of affection, incorporating a symbolic rating system that reflects their personal perceptions or biases toward different nations. Such rankings, while seemingly playful or informal, may carry deeper cultural or geopolitical connotations, shaping how users interpret and engage with the post. This interaction style highlights the complex interplay between subjective evaluation and social dynamics within the digital soundscape, where personal opinions can influence collective discussions and reinforce existing stereotypes or narratives.

The author's evolving attitude toward Turkey is also worthy of discussion. The first mention of Turkey appeared in a comment dated September 3, 2024, where the author stated, "We are not brothers. Japanese people normally would put 🟥. I like Turkey but I'm kinda alone on this one in my country 😅." This initial comment reflects a personal and somewhat isolated appreciation for Turkey, suggesting a divergence from the general sentiment within the author's home country. However, as the number of comments from Turkish users increased, the author’s tone and engagement appeared to shift, potentially indicating a growing sense of camaraderie and friendly interaction. On November 23, 2024, the author revised their comment to, "Before it was 🟩, now 🟦 :)," marking a clear change in the rating of Turkey from a neutral stance (🟩) to a more positive one (🟦).This shift in the author's comments is significant, as it may reflect the influence of social interactions and user engagement within the digital space. The increasing participation from Turkish users seems to have fostered a more positive view of Turkey, suggesting that online interactions, in which users share their perspectives and engage with one another, can have a tangible impact on the author’s subjective evaluation. This change underscores the dynamic nature of digital discourse, where public sentiment can be shaped and modified over time through continuous engagement and evolving social connections. 6

Soundspace Analysis
The background music used in the video is an excerpt from the French song "L'Amour Toujours (Speed Up)" . While the song is not a traditional French piece, it gained widespread popularity across Europe, including France, due to its significant influence and global reach. The Speed Up version of the track amplifies its dynamic qualities, making it more suitable for fast-paced environments, and contributing to its widespread usage on TikTok. As of now, the audio has been used in 15.8K videos.Given the large number of videos associated with this audio, we categorized the types of videos involved. Due to time constraints, we were unable to analyze all the videos, so we focused on the first 600 videos. From this subset, we identified three predominant categories: political, religious, and entertainment content. Within these categories, we concentrated our analysis on the political group, specifically the videos with the highest levels of comment and like interactions. This approach allowed us to gain insights into how the background music and its associated commentary fostered specific discursive environments, particularly within the context of political engagement on TikTok.

Snippet from poster #1, link on top of page.

5. Conclusions

The political campaign of the Alternative für Deutschland (AfD) has been very effective in taking advantage of TikTok ’s multimodal affordances. PR-tricks like letting a party leader arrive at an event on a German-signature motorbike, or associating your ideologies with an upbeat clubby song like Gigi d’Agostino’s l’Amour Toujours, paid off wildly as these content strategies reverberate across modalities and publics. Participatory practices channel and prolong far right political ideologies through riffing on popular songs, meme templates, and visual symbols that are seeded on the platform.

6. References

Christin, A., Bernstein, M. S., Hancock, J. T., Jia, C., Mado, M. N., Tsai, J. L., & Xu, C. (2024). Internal Fractures: The Competing Logics of Social Media Platforms. Social Media+ Society, 10(3), 20563051241274668.

Duguay, S., & Gold-Apel, H. (2023). Stumbling blocks and alternative paths: reconsidering the walkthrough method for analyzing apps. Social Media+ Society, 9(1), 20563051231158822.

Geboers, M., & Pilipets, E. (2024). Networked masterplots: Music, pro-Russian sentiment, and participatory propaganda on TikTok. Journal of Digital Social Research, 6(1), 90-103.

Hee et al. (2024). Recent Advances in Hate Speech Moderation: Multimodality and the Role of Large Models. Available at: https://arxiv.org/abs/2401.16727

Medina Serrano, J. C. (2021). Multiplatform Analysis of Political Communication on Social Media (Doctoral dissertation, Technische Universität München).

Pilipets, E., Geboers, M. & Delavar-Kasmaii, D. (2025). Detouring, Rerouting, Weaponization: Memetic Soundscapes and the Secondary Orality of WarTok. In: Mutsvairo, B, Nguyen, D., Zeng, J. (eds). Technology, Power & Society Global Perspectives on the Digital Transformation, Brill Books

Steen, E., Yurechko, K., & Klug, D. (2023). You can (not) say what you want: Using algospeak to contest and evade algorithmic content moderation on TikTok. Social Media+ Society, 9(3), 20563051231194586.