Curated politics:

A study of Black Lives Matter protests on TikTok using the digital snowball method

Team Members

Shuaishuai Wang | Jeroen de Vos | Daniela Jaramillo-Dent | Valeria Donato | Yan Asadchy | Caio Mello | Andrea Elena Febres Medina| Teresa Cremonesi | Jason Armitage | Golsa Tahmasebzadeh | Endri Kacupaj | Swati | Anna Jørgensen


Team Members 1

Contents 1

Summary of Key Findings 2

1. Introduction 2

2. Initial Data Sets 3

3. Research Questions 3

4. Methodology 3

1) API retrieval of the data 4

2) video categories & creative techniques 4

3) the role of music: distribution and engagement 5

4) symbolic practices 6

5. Findings 6

5.1 Video categories and creative techniques 7

5.2 Creative techniques and engagement 8

5.3 Audio files and engagement 11

5.4 Symbols of the Black Lives Matter movement: the fist 13

6. Discussion 13

7. Limitations 14

8. Conclusion 15

9. References 15

Summary of Key Findings

Political content has been taking shape on TikTok on its own terms, which is profoundly shaped by the “For you” algorithms.
  • There is a significant discrepancy between the most recommended and the most engaged videos. Peaceful and creative videos receive more views, while provocative and violent videos provoke more engagement.
  • Even using a variety of creative techniques, provocative videos receive more comments but do not appear among the most viewed ones.

1. Introduction

Often self-positioning as a platform for fun and creativity, TikTok has been keenly isolating itself from politics ever since it achieved popularity among young audiences around the globe. According to documents leaked to the media, TikTok labels political content relating to demonstrations and controversial topics (e.g. separatism, and religion, racial and ethnic conflicts) as “not recommend” or “nor for feed” (Köver and Reuter, 2019) As such, users are restricted from political exposure on the platform, even if it has abundance of political content. For this reason, TikTok ’s recommendation algorithms have received a torrent of criticism.

In practice, these "not recommended" videos are not removed, but are tactfully hidden from users’ algorithmically fed “For You” page after they reach a limited view count. This biased algorithmic practice suggests that TikTok manipulates content visibility in ways of drowning out or “deplatforming” voices of certain groups, events, or issues(Rogers, 2020). Content categorized as “not recommend” and “not for feed” on TikTok is reportedly being regarded as politically problematic or posing bullying risks to vulnerable groups (Köver and Reuter, 2019). Moreover, TikTok ’s algorithmic design is based on users’ interaction history. Given the timeliness of political topics and their emergent nature, it provides the platform an perfect excuse for hiding controversial politics from users whose past interaction history does not contain political content.

To understand how political content takes shape in the biased algorithmic system of TikTok, this project proposes a new methodological approach to the study of political videos of TikTok using recommendation algorithms—what we call the digital snowball sampling method. Networking and referral are central to the traditional snowball sampling method: a small number of initial contacts (seeds) recommend people akin to them to a study (Parker, Scott and Geddes, 2019). This method has been widely used to locate hidden and stigmatized populations such as LGBT and immigrants (Atkinson and Flint, 2001; Browne, 2005). In what has been termed “virtual snowball sampling,” researchers have successfully harnessed the networking capacity of Facebook Groups to recruit hard-to-reach immigrants online (Baltar and Brunet, 2012). These online referrals further led the researchers to more community members in reality.

“Digital snowball” differs from the “virtual snowball” in three aspects. First, social media platforms are not used to locate respondents in the offline world. Instead, it specifically focuses on people’s online activities and how they produce content for the internet. Second, recommendation algorithms, which operate as “method of the medium” (Rogers, 2013), act as the proxy to conduct the sampling process on behalf of researchers. Unlike “virtual snowball” that is adapted for locating people in the physical world, digital snowball originates from a “natively digital” logic (Rogers, 2013). As such, it explores culture not simply made on the internet, but more by means of it (Rogers, 2013). Third, besides sampling hard-to-reach populations, the project reinvents the snowball method as a generic approach to the study of the messy and intertwined internet phenomena. More specifically, the digital snowball method can be used for sampling videos of related kinds (e.g. hate speech, cyberbullying, sexism, and racism) on auto-playing short video platforms represented by TikTok.

In this project, we apply the digital snowball method to sampling short videos that are related to Black Lives Matter protests on TikTok in order to investigate how the recommendation algorithm shapes and engages with political contents on the platform.

The project aims to surface how the “For You” page recommends certain types of blm’s content, more peaceful and creative, although others get more engagement.

2. Initial Data Sets

200 videos are sampled (initial dataset) in a research account previous to the summer school. The TikTok research account was accessed in Amsterdam, with the liked videos being in English.

3. Research Questions

The research builds on top of two separate research questions, concerning curated politics in algorithmic governance, and appropriation of creative affordances.
  1. How do recommendation algorithms shape the political landscape of TikTok?
  2. How do TikTok ’s creative affordances affect user engagement of Black Lives Matter protests?

4. Methodology

In preparation for the DMI summer school, we created a researcher’s profile without any additional interaction on 3 June 2020. TikToK had received media criticism for hiding the Black Live Matters protests-related videos (Harris, 2020). In response, the platform became proactive in recommending Black Live Matter movement videos to users. For this reason, the research account encountered the first Black Live Matter video within 5 minutes of scrolling without identifying the seed accounts to follow. As the research account liked the first video concerning the #blacklivesmatter topic, we found that the algorithm would indeed feed us more relevant #blacklives matter related videos. Rather than filtering or selecting manually, we used this form of digital snowball sampling by relying on Tiktok’s personal recommendation algorithm feeding us more relevant content.

As a consequence, between June 3rd and July 30th, we sampled 200 videos from the manual liking activity via the research account. In the sampling process, we identity the videos based on the hashtags (#blacklivesmatter or #blm) used and the visuals (if the content contains protest information). It needs to be noted that our ‘preference’ for #blacklivesmatter content is only one of the data points used by TikTok ’s algorithm to determine relevant content. As with many social media companies user profile is also derived from data not explicitly shared by the user (eg. IP address, type of devices, browser language preference etc.), Tiktok has been criticized for collecting large amounts of data - prompted by some as modern spyware.

In our subsequent analysis we took three different steps 1) API retrieval of the data 2) close reading of the videos developing and employing two categorisation protocols and 3) looking into distribution and engagement around the music used in the clips and 4) we looked into symbolic practices tagging the use of specific Black Lives Matter symbols.

1) API retrieval of the data

TikTok provides a vast range of ways to discover and share the contact: through hashtag(challenge in the metadata), music used in video and/or through direct search query. For the purpose of this study, we used found videos of interest through the hashtag #blm and #blacklivesmatter and liked a sequence of these videos to build up the profile of interests for the recommendation engine that will define the “For You” page and suggest related content. As the next step we liked 200 suggested videos related to the #blm movement. These videos formed a corpus of data for scraping.

To scrape the data, we used an API Wrapper for Python by Avilash Kumar (

This wrapper allowed us to extract all the available metadata from the liked videos and use it for the further analysis. The Python Code used is provided in the Appendix 1.(if needed). The result of the scraping and cleanup is a csv file with the metadata that includes unique hashed ID, sound type, amount of views, likes, shares, comments.

2) video categories & creative techniques

Through collectively reviewing a set of sample videos, we identified a number of video categories and creative techniques (or affordances) in TikTok video creation. We subsequently created a shared glossary of terms to ensure the conformity of video analysis (see tables 1 and 2). We subsequently categorized every video manually, ‘close reading’ the 200 videos in our sample.


Table 1: video categories


Table 2: creative techniques

3) the role of music: distribution and engagement

TikTok creates new possibilities for engaging with the content and audiences through its design. Music as an entry point and connection link for opinions.

In the context of this study, we explore how communication trees appear as users create new content as a response to existing one. Usage of music and original sound provide new ways of discoverability, recreation and dissemination of the information across users of TikTok and other platforms. Considering the important role of social media in delivering political news updates to a large portion of the population, these multi-level distribution channels become a powerful tool for engagement.

Until recently, social media platforms like Facebook, Twitter and YouTube were considered the most relevant mediators of political content (Tucker & Guess, 2018). However taking into account the dynamic nature of social media, it is possible for other platforms to become more popular (Lustig & Pine, 2016). As TikTok recently became the second most downloaded application in 2019, there is a need to pay closer attention to it’s design and how it facilitates the political dialog (Williams, 2020).

Current study aims to understand what role music plays in the #blacklivesmatter movement on TikTok and specifically to understand how differently users of TikTok engage with the content that features original sound(e.g. personal comments or ambient sound of demonstration) and songs of popular artists. Understanding how the design of the platform and its features will bring light over the role of music and how it shapes the way political information can be reused and spreaded by users or limited by the platform.

4) symbolic practices

In the process of coding, we identified symbols that were prevalent throughout the sample, in specific the clenched fist and other variations of this symbol. We decided to do one more round of coding to account for the presence of the fist and the different modalities/formats in which it appears.

The use of affordances specific to this platform that resemble the vernacular used by mainstream TikTokers is also a symbolic consideration that was coded within codes such as “monologue”, “lipsync”, “green screen”, among others. These become symbolic because they attempt to align with other uses of the platform while advancing the #BlackLivesMatter movement message.

Other symbolic practices involve the performative nature of TikTok videos, represented by the use of the body to symbolize the violence, discrimination, and empowerment of the #BlackLivesMatter movement. Creative practices such as dancing and music were also present but were coded within the “creative practices” code without specifying the specificities which will be done in subsequent studies.

5. Findings

5.1 Video categories and creative techniques


Figure 1. Overview of c reative techniques and content categories

The most common video category was labeled as “peaceful” followed by videos labeled as “provocative” and “violent” which were present in similar numbers. The least common videos in our sample include “sarcastic” and “emotional” videos.

In terms of the creative techniques used in each category, it is worth noting that “word bubbles” and “sound” were the most prevalent across all video categories. Moreover, the “peaceful videos” did not include creative techniques such as “green screen”, “lip-sync”, “duet”, “facial expression” or “reaction video”.

On the other hand, “informative”, “emotional” and “sarcastic” videos do not feature “dance”, “visual effects” or “remix” as creative techniques.

Finally, the “monologue” format which is one of the characteristic video types on the platform, is only present in “provocative” and “informative” videos, alongside the “facial expression” label, which is also present in the same video classifications.

5.2 Creative techniques and engagement


Figure 2. Most viewed videos in terms of video categories and creative techniques

In our sample, the most viewed videos are “creative” and “peaceful” with a diverse array of creative techniques. Furthermore, the most prevalent creative technique is “sound”. In this case, it is interesting to note that only four types of videos emerge as most viewed, namely “creative”, “emotional”, “peaceful”, and “violent” excluding all other types of videos in the sample.


Figure 3. Most liked videos in terms of video categories and creative techniques

Conversely, in the next level of engagement, “violent” and “peaceful” videos lead in terms of number of likes with “sound” and “word bubbles” as the main creative techniques (Figure 3)


Figure 4. Most shared videos in terms of video categories and creative techniques

In terms of the most shared videos, those labeled as “provocative” and “violent” are among the most shared and “political speech”, “sound” “word bubbles” are the leading creative techniques used within this type of interaction from the user (Figure 4).


Figure 5. Most commented videos in terms of video categories and creative techniques

Most commented videos are similar to most shared, as those videos labeled as “provocative” and “violent” videos prevail, although in this case “sound” and “word bubbles” are the most common creative techniques and “political speech” is not as prevalent, in contrast with most shared videos.

In this case, provocative videos that received the most comments in the sample display a wide array of creative techniques.

5.3 Audio files and engagement

The “discoverability through sound” functionality that characterizes TikTok from other platforms is relevant to the present analysis as it enables users to appropriate and “recycle” a previously used audio file for their new creation. This makes TikTok a unique platform, where lip syncing, social media challenges and a clear sonorous identity are a relevant part of its vernacular.

Out of 200 sampled videos, 123 had original sound incorporated in them, making it a 61,5% of the sample. The audio used in the video with its relation to the whole sample under study is presented on Figure 6.


Figure 6. Audio tracks used in the sample

In this sense, and within our sample, the most common audio type is what the platform calls “original sound” (most used audio tracks among the sample of 200 videos) which is a sound recorded by the creator, and may include speech, ambient sounds and in many cases is a music track. However, this type of audio is not among the most engaged-with in terms of views, comments and shares as we can see in the following graph, presented on the Figure 7 below.


Figure 7. Most liked videos in terms of video categories and creative techniques

In this graph, it is evident that videos with the highest average play count (represented by the size of the circles) reflects that, although most videos in the sample feature “original sound” , this is not mean a great discoverability and engagement in sense average play and share counts, with five popular songs leading and suggesting a sort of “soundtrack” of the movement. These include, in terms of average play count: “Coño (feat.Jhorrmountain x Adje) Puri” and “Childish Gambino -This is America/Post Malone- COngratulations carneyval_”. In terms of average share count, the most successful audio tracks are “Respect Aretha Franklin”, “Immortal Beloved (instrumental) BLUKSHP” and “Childish Gambino -This is America/Post Malone- Congratulations carneyval_”. These numbers suggest that a large portion of engagement and discoverability comes from the popularity of the song incorporated.

5.4 Sym bols of the Black Lives Matter movement: the fist

The clenched fist is one of the identifying visual traits of the Black Lives Matter movement, both online and offline. In our sample of 200 short videos, the clenched fist appears in more than half (n=107) of the videos. Figure 8 shows the different ways in which this symbol appears within the sample analyzed.


Figure 8. The different ways in which the fist appears in the #BLM movement TikTok short videos.

The presence of this symbol, together with the prevalence of certain music tracks, hashtags, performative practices, along with other traits arguably build a powerful identity for the #BLM movement on TikTok.

6. Discussion

Results suggest that TikTok ’s policies for content moderation and the algorithmic ”For You” page follow specific guidelines in line with information leaked and reported by Netzpolitik (Köver & Markus, 201 9) and other media. For instance, in this sample, videos coded as “peaceful” and “creative” displayed the highest number of views, which could suggest that the algorithm is more likely to promote this type of content as opposed to “political speech” in line with the aforementioned leaked moderation guidelines (Hern, 2019a; Chen, 2019).

Hearn (2019a) found that TikTok limits its political content in electoral periods. In this sense, some of the content moderation rules of interest for this sample include the labeling as “not for feed” and “not recommended” of content related to police and riots/protests (Chen, 2019). This may have motivated the initial blocking of the hashtags #BlackLivesMatter and #GeorgeFloyd, for which TikTok later apologized (Harris, 2020). Although the glitch that caused these hashtags to be blocked, our analysis provides a glimpse of what these algorithmic policies may imply for political content and content related to activism.

The use of certain music tracks and sounds is also interesting due to the prevalence of certain music within the sample also suggests that the audio used may have some effect in whether it appears in the feed or not. In our case, although original sound was the most prevalent but it was not among the most played or shared suggesting some relationship between using original sound and becoming highly visible in the “for you” feed.

Another interesting finding involves a sort of “branding” of the movement consisting of performative acts within creative practices, the fist as a symbol and the prevalence of certain music tracks that become the “soundtrack” of the BLM movement. This is an aspect that we will expand in future research collaborations.

7. Limitations

A limitation that came up during our discussions was the difficulty to assess videos on the basis of traditional metrics of success (views, likes, comments and shares) when some of them were posted earlier than others, giving them more time to acquire likes, shares and views. This is an important consideration, since popular and viral content needs some time on the platform, and in this case the sampling had been done throughout one month and included one-month-old videos alongside one-day-old ones.

An ethical issue arose during the data collection process. As 60% of American users of TikTok, who are posting and engaging with #blm content are between 16 to 24 years old (Wang, 2019), they might not fully realise that their data is public and exposed to researchers. To maintain the data protection, all the names were deleted prior classifying the videos and analysis of songs used.

Future research on TikTok could delve into the significance of sound and effects as functionalities that allow content to be found, replicated and reinvented in the platform. A deeper understanding of these affordances would also allow for a deeper understanding of algorithmic moderation, which may also be done on the basis of these specific and unique configurations. The effect of these on TikTok ’s vernacular are also of interest, since it shapes the creativity of users and it promotes the amplification of certain types of contents.

Many countries are currently blocking TikTok, which has important implications for the platform and also for the content creators, research delving into the motivations and implications of this could be illuminating.

8. Conclusion

Political content has been taking shape on TikTok on its own terms, which is profoundly shaped by the “For you” algorithms.
  • There is a significant discrepancy between the most recommended and the most engaged videos. Peaceful and creative videos receive more views, while provocative and violent videos provoke more engagement.
  • Even using a variety of creative techniques, provocative videos receive more comments but do not appear among the most viewed ones.

In conclusion, we have seen that TikTok has the potential to impact on political communication and engagement, in particular from teenagers perspective. Aware of the fact that the platform allows users to register from the age of 13, this has prompted an ethical consideration on the treatment of analyzed data. In fact, given their young age, users may not be fully aware of publishing their features and behaviors on the platform. This concern must be kept in mind when researchers collect and manage data from TikTok.

9. References

Atkinson R and Flint J (2001) Accessing hidden and hard-to-reach populations: snowball research strategies. Social Research Update. Retrieved from:

Browne K (2005) Snowball sampling: Using social networks to research non-heterosexual women. International Journal of Social Research Methodology 8(1): 47–60.

Baltar F and Brunet I (2012) Social research 2.0: Virtual snowball sampling method using Facebook. Internet Research 22(1): 57–74.

Chen A (25 November, 2019) A leaked excerpt of TikTok moderation rules shows how political content gets buried. MIT Technology Review. Retrieved from

Harris M (1 June, 2020) TikTok apologized for the glitch affecting the 'black lives matter' hashtag after accusations of censorship: 'We know this came at a painful time'. Insider. Retrieved from

Hern A (25 September, 2019a) Revealed: how TikTok censors videos that do not please Beijing. The Guardian. Retrieved from

Hern A (3 December, 201 9b) TikTok owns up to censoring some users' videos to stop bullying. The Guardian. Retrieved from

Köver C & Markus R (2 December, 201 9) TikTok curbed reach for people with disabilities. Netpolitik. Retrieved from

Lustig C, Pine K, Nardi B, Irani L, Kyung Lee M, Nafus D & Sandvig C (2016) Algorithmic authority: the ethics, politics, and economics of algorithms that interpret, decide, and manage. In Proceedings of the 2016 CHI Conference Extended Abstracts on Human Factors in Computing Systems. ACM, 1057–1062.

Parker C Scott S and Geddes A (2019) Snowball sampling. In: Atkinson P, Delamont S, Cernat A, Sakshaug JW and Williams RA (Eds) SAGE Research Methods Foundations.

Rogers R (2013) Digital Methods. Cambridge, MA: The MIT Press.

Rogers R (2020) Deplatforming: Following extreme Internet celebrities to Telegram and alternative social media. European Journal of Communication, 35 (3), 213-229.

Serrano J C M, Papakyriakopoulos O & Hegelich S (2020) Dancing to the Partisan Beat: A First Analysis of Political Communication on TikTok. In 12th ACM Conference on Web Science (pp. 257-266). Retrieved from

Tucker JA, Guess A, Barberá P, Vaccari C, Siegel A, Sanovich, S, Stukal D & Nyhan B (2018) Social media, political polarization, and political disinformation: A review of the scientific literature [Report]. William and Flora Hewlett Foundation. Retrieved from

Wang E, Alper A, Roumeliotis G & Yang Y (1 November, 2019) U.S. opens national security investigation into TikTok - sources. Reuters. Retrieved from

Williams,K. (26 January, 2020) TikTok Was Installed More Than 738 Million Times in 2019, 44% of Its All-Time Downloads. Sensor Tower. Retrieved from

Appendix 1

import http.client

import re

import csv

from TikTokAPI import TikTokAPI

## Create an object

api = TikTokAPI ()

## Collect all the videos liked by user and store them in a collection

user_videos = api.getLikesByUserName("jk115955", count=1)


## Create a list that stores only entities of a collections with the key "Items"

list = user_videos.pop('items')

## Print the content of the first item of the list:


## Used to check type of the API Call output, as GitHub repository didn't provide


## Variables for the While loop

i = 0

k = 0

l = 0

## Extracting relevant information

while i < 1:


i += 1

## Printing into a CSV file

## Define columns

columns = ['id', 'desc', 'createTime', 'video', 'author', 'music', 'challenges', 'stats', 'originalItem',

'officalItem', 'textExtra', 'secret', 'forFriend', 'digged', 'itemCommentStatus', 'showNotPass', 'vl1']

## Name a file

csv_file = "TikTokScrapedData.csv"

## Write to the file


with open(csv_file, 'w', newline='', encoding="UTF-8") as csvfile:

writer = csv.DictWriter(csvfile, fieldnames=columns)


for data in list[0]:


except IOError:

print("IO error. Please check the input data")

-- JedeVo - 22 Jul 2020
Topic revision: r2 - 22 Jul 2020, JedeVo
This site is powered by FoswikiCopyright © by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding Foswiki? Send feedback