By Anna Shams Ili, Junior Researcher


Few social media platforms are known for their excellent API access, but TikTok more than most is keen on gatekeeping any access to their data, especially when it comes to data around the recommender algorithm. From a business standpoint, this is understandable, as the research that has been produced around TikTok has often been less favourable. Just last week (at the time of writing), the Danish organisation and analytics office Digitalt Ansvar released a report on the detrimental effects of TikTok and how the algorithm leads vulnerable users to more extreme content.

Digitalt Ansvar conducted an experiment where they were able to record how profiles playing as young people who liked content around mental health and depression were led into more extreme content encouraging self-harm and other “psychologically damaging” content. For anyone who has researched or is simply interested in social media algorithms, this pattern will sound all too familiar, but the report highlights how disturbing content can not only thrive on TikTok but be recommended to the platform’s most vulnerable users.


TikTok and algorithms

TikTok’s primary function is its “For You” feed – a continuous feed of short-form videos recommended to you by the algorithm. According to TikTok itself, the company uses user interactions (liking, sharing, commenting), video information (captions, hashtags, and so on), and device and account settings (such as your location) to recommend you videos. The social media platform became big fast, and the way of presenting content is so new that people still frequently associate the platform with their 12-year-old niece, although the main demographic is said to be in the 18-35 age range. Internal company documents, when leaked, become explainers in The New York Times or exposés in the Guardian, where it in 2020 was revealed that the platform tried to exclude content creators who were “abnormally shaped” or in an environment that was “shabby and dilapidated.” Everyone wants to know what is in their secret sauce, but it is simultaneously clear that whatever it is, it’s probably not great.

Algorithmic folklore or folk theories are used to describe how users conceptualise algorithms and are primarily used to understand how users shape their behaviour according to this understanding. Returning to the Digitalt Ansvar report – in their findings, most content that is labelled as disturbing or promoting self-harm does not directly use the terms self-harm or contain visuals of self-harming. Instead, users use euphemisms, in-jokes, or terms such as “sewercide” and “unalive”. The latter two terms are both types of “algo-speak” used to refer to suicide. Algo-speak is used by users to circumvent moderation and has become a common practice on the platform. Simultaneously, it makes it harder to study content systematically, as words and meanings may be obscured.


Everything is “girl”, online

If you’ve been paying attention to the internet, you know everything is girlie – bows on everything, girl math, bimbo-feminism, and girl dinner are just some of the few trends that have been circulating on my feed the past year. Topics range from empowering reclamations of girlhood or femininity to “explaining geopolitics to girls” through shopping metaphors. Of course, it didn’t start this past year – Teen Vogue in 2021 already wrote on the traditional gender norms revived through “StraightTok”, and the romanticisation of the “TradWife” was seen as connected to post-COVID burnout for many. But for many, 2023 was when “girl” became inescapable on the app.

Around late spring or early summer (coincidentally overlapping with two other TikTok phenomena: Barbie and the Taylor Swift eras tour), the TikTok trend “girl dinner” popped up. As with many TikTok trends, the original meaning is in the eyes of the viewer. Evil tongues, mine included, may claim that this was yet another “girl” moniker denoting an innate connection between femininity and the object of the video – in this case, eating patterns. Others claimed the “girl” part was a tongue-in-cheek call to fellow “lazy girls”, who couldn’t bother to cook. As someone who has never enjoyed cooking, I am partial to olives and cheese as a meal as well, although I don’t view it as particularly gender-defining. The trend and surrounding discussion spurred my thoughts around TikTok’s obsession with girlhood. This girlhood was almost always used as referring to a “universal”, yet primarily white, straight, and middle-class, experience.

What especially struck me was that within all these different forms of traditional femininity or aesthetic acclamations of what being a girl constitutes, was at a core still the concept of one specific type of beauty, and particularly one type of body. The girl dinner trend seemed benign in itself, but the connotation of it being a “girl thing” that was specifically feminine, reminded me of yet another trend: “coquette-core”, an aesthetic that is inspired by vintage fashion and the idea of being “delicate”, almost Victorian-doll-like. If you were on Tumblr during the 2010s, it was an aesthetic also often closely associated with eating disorders and “thinspiration” – a term used widely in the pro-ana community.

Pro-ana stands for “pro-anorexia”, and refers to online communities that support each other in their eating disorders. Support in this sense can refer to both sharing stories of recovery, however also supporting in sharing tips on how to maintain and hide eating disorders. Online pro-ana communities have existed since messaging boards, but TikTok is a new avenue for this content. In a 2020 article on pro-ana content on TikTok, Wired wrote: “Without dismissing anyone’s claims about their For You recommendations, readers should know that users who are not engaging with videos related to eating disorders are highly unlikely to have them randomly recommended.” TikTok has since also banned hashtags such as #anorexia or #proana, but the content still exists, and it is being recommended to someone. So, who?


From networked conspiracy to actual networks

At this point in my stream of consciousness, I felt like a conspiracy theorist.

And like any good conspiracy theorist, I’ve always enjoyed linking my ideas in a mindmap. So it’s maybe not strange, that I’ve always found networks to be interesting, especially when looking at networks of ideas. I started off this post by saying how difficult TikTok is to study. Another reason why it is difficult is the search function. TikTok doesn’t simply search for a term based on actual text but uses fuzzy name-matching, a way of determining what words are similar to the search term. Unfortunately, it is not the most accurate. In one paper, the search result for the defense alliance NATO returns videos related to the anime “Naruto” instead – a close match in name, but not in meaning[i]. Beyond the name-matching, the paper raises multiple methodological issues with studying TikTok: the search result also doesn’t return everything on the platform. I recently experienced this myself when I scraped data for an assignment and found that over 50% of my results were from the past 3 months. TikTok search results are biased towards both recency and popularity.

To transform the loose ideas of trends and concepts into something tangible, and keeping in mind TikTok’s limitations, hashtags seemed like the most obvious answer. Users use specific hashtags to reach audiences, and data can be double-filtered to exclude rows that don’t contain a match, thereby avoiding the Naruto-problem. When writing this post, I was trying to find my old notes of my idea, unfortunately, they are not as illustrative as I had hoped.


Luckily, I got to unfold the conceptual idea further at the Play Lab. Participants were instructed to pick a starting point, and then connect different hashtags based on what they thought would fit.


Using Kozinet’s idea of netnography and immersion, I have set up a blank TikTok profile, as I, unfortunately, don’t have the same means as Digitalt Ansvar to get a completely clean phone. As a chronically online student, I had starting points though, and started collecting observations on the platform and finding pro-ana content. This is not going to be a guide on how to do that, but I was surprised at how easy it was. I created my initial seed list for scraping and have had a preliminary look at co-occurring hashtags, including content around eating disorder recovery and K-Pop idols. My next steps are a lot of scraping, and then selecting some of the most relevant co-occurring hashtags to repeat the process. As the idea-generated networks in the play lab, there are multiple ways down the rabbit hole, and I hope that this project can shed some light on at least one of them.   


[i] Steel et al. (2023): The Invasion of Ukraine Viewed through TikTok: A Dataset