Product Code Database
In social media, algospeak is a self-censorship phenomenon in which users adopt coded expressions to evade automated content moderation. It allows users to discuss topics deemed sensitive to moderation algorithms while avoiding penalties such as shadow banning, downranking, or de-monetization of content. A type of netspeak, algospeak primarily serves to bypass censorship, though it can also reinforce group belonging, especially in marginalized communities. Algospeak has been identified as one source of linguistic change in the modern era, with some terms spreading into everyday offline speech and writing.
Between July and September 2024, TikTok reported removing 150 million videos, 120 million of which were flagged by automated systems. Automated moderation may miss important context; for example, benign communities who aid people who struggle with self-harm, suicidal thoughts, or past sexual violence may inadvertently receive unwarranted penalties. TikTok users have used algospeak to discuss and provide support to those who self-harm. An interview with nineteen TikTok creators revealed that they felt TikTok's moderation lacked contextual understanding, appeared random, was often inaccurate, and exhibited bias against marginalized communities.
Algospeak is also used in communities promoting harmful behaviors. Anti-vaccination Facebook groups began renaming themselves to “dance party” or “dinner party” to avoid being flagged for misinformation. Likewise, communities Pro-ana the eating disorder anorexia nervosa have been employing algospeak. Euphemisms like "cheese pizza" and "touch the ceiling" are used to promote child sexual abuse material (CSAM).
On TikTok, moderation decisions can result in consequences such as account bans and deletion or delisting of videos from the main video discovery page, called the "For You" page. In response, a TikTok spokeswoman told The New York Times that the users' fears are misplaced, saying that many popular videos discuss sex-adjacent topics.
Techniques used in algospeak are extremely diverse. In orthography, users may draw from leetspeak, where letters are replaced with lookalike characters (e.g. $3X for sex). Certain words or names may be censored, or in the case of auditory media, cut off or Bleep censor, e.g., s*icide instead of suicide. Another involves "pseudo-substitution", where an item is censored in one form, while it is present in another form at the same time, as used in videos. Some may involve intersemiotic translation, where non-linguistic signs are interpreted linguistically, in addition to further obfuscation. For example, the corn emoji "" signifies pornography by means of porn→ corn→🌽. Others may rely on phonological similarity or variation, such as homophobic→ hydrophobic, and sexy→ seggsy via intervocalic voicing. On Chinese social media, users sometimes replace sensitive terms with characters that differ only in tone. For example, 细颈瓶 (xì jǐng píng, literally “narrow-necked bottle”) is used as a stand-in for the name of General Secretary of the Chinese Communist Party, Xí Jìnpíng.
In an interview study, most creators that were interviewed suspected TikTok's automated moderation was scanning the audio as well, leading them to also use algospeak terms in speech. Some also label sensitive images with innocuous captions using algospeak, such as captioning a scantily-dressed body as "fake body". The use of gestures and Emoji are common in algospeak, showing that it is not limited to written communication. (2025). 9781450394192 ISBN 9781450394192 A notable example is the use of the watermelon emoji on social media as a pro-Palestine symbol in place of the Palestinian flag in order to avoid censorship by Facebook and Instagram. Black creators may simply present their light-colored palms to the camera to stand in for white people, and flip them to stand in for black people.
Algospeak can lead to misunderstandings. A high-profile incident occurred when American actress Julia Fox made a seemingly unsympathetic comment on a TikTok post mentioning "mascara", not knowing its obfuscated meaning of sexual assault. Fox later apologized for her comment. In an interview study, creators shared that the evolving nature of content moderation pressures them to constantly innovate their use of algospeak, which makes them feel less authentic.
A 2024 study showed that GPT-4, a large language model, can often identify and decipher algospeak, especially with example sentences. Another study shows that sentiment analysis models often rate negative comments incorporating simple letter–number substitution and extraneous hyphenation more positively. (2024). 9798350351309, IEEE. EAN 9798350351309
|
|
