# Speaking in Code: The Hidden Language Born from Algorithm Fear
The word "mascara" started appearing in comment sections around 2021. Not in beauty tutorials. In trauma disclosures. Survivors of sexual abuse had quietly agreed, somewhere in the distributed consciousness of TikTok, that "mascara" would stand in for the thing the algorithm wouldn't let them name. The substitution spread fast, person to person, video to video, until it was everywhere — a secret that wasn't quite a secret anymore. Then actress Julia Fox saw the word and responded warmly, cheerfully, about actual makeup. She apologized when someone explained what she'd missed. But that moment of collision — between the coded world and the uncoded one — revealed something strange had already happened. A shadow language had grown up inside the internet, and not everyone could read it.
Fifty million videos. That's roughly how many TikTok removes in a single month. Between July and September 2024, the platform reported pulling 150 million videos total, 120 million of them flagged before a single human ever saw them. Automated systems. Pattern-matching at a scale no moderation team could replicate. The machine scans, the machine decides, and the content disappears — sometimes correctly, sometimes not, and often without explanation.
The language that emerged in response has a name now: algospeak. A portmanteau of "algorithm" and the Orwellian suffix "-speak," the term appears to have surfaced around 2021, though the behaviors it describes are older. Internet researcher Emily van der Nagel was tracking the phenomenon as early as 2018, when she coined the terms "Voldemorting" — refusing to name something directly, like the villain whose name must not be spoken — and "screenshotting," the practice of posting text as an image to dodge keyword filters. These were early, clumsy workarounds. What came after was something more systematic. Something almost alive.
To understand why algospeak spread so fast, you have to understand the specific anxiety of the creator economy circa 2020 and 2021. TikTok had become a genuine livelihood for hundreds of thousands of people. Reach was everything. A single video could generate income, sponsorships, a career — but only if the algorithm surfaced it. And the algorithm, as creators experienced it, was a black box with a hair trigger. Words related to sex, death, drugs, self-harm, political controversy: all of them seemed to suppress a video's reach, sometimes burying it entirely in a process creators called being "shadowbanned." TikTok's own spokeswoman told The New York Times that fears about moderation of sex-adjacent topics were misplaced. Creators, almost universally, disagreed. The gap between the platform's official position and the lived experience of its users was where algospeak was born.
The substitutions multiplied in every direction. "Corn" for pornography — the chain running from "porn" to "corn" to the corn emoji, each step a degree further from the original. "Unalive" for suicide or death, a clinical-sounding evasion that spread so widely it now appears in mental health discussions with no irony at all. The watermelon emoji as a symbol of Palestinian solidarity on Facebook and Instagram, where the Palestinian flag itself seemed to trigger suppression. Anti-vaccination Facebook groups renaming themselves "dance party" and "dinner party" to slip past misinformation flags. On Chinese social media, users replaced politically sensitive terms with tonal homophones — 细颈瓶 standing in for Xi Jinping's name, the substitution legible to any Chinese speaker and invisible to a keyword filter. Communities promoting anorexia nervosa developed their own coded vocabulary. So did people discussing drug use, sex work, and chronic illness. The language fractured into a dozen dialects, each one a map of what a specific community feared losing.
Linguists began cataloguing the techniques. Leetspeak — swapping letters for numbers, writing "s3x" instead of "sex." Emoji substitution. Phonological similarity, where a word sounds like the forbidden term without spelling it. Intersemiotic translation, moving meaning from text into image or symbol. Pseudo-substitution, where a word is replaced by something tangentially related but not phonetically similar — "mascara" for sexual abuse, the connection legible only to insiders. Each technique exploited a different gap in how automated systems process language. Collectively, they constituted a living linguistic arms race.
Taylor Lorenz brought the term "algospeak" to mainstream attention in a 2022 Washington Post article, and that same year a poll found that nearly a third of American social media users reported using emojis or alternative phrases specifically to subvert content moderation. A third. That's not a subculture. That's a mass behavior, operating mostly beneath the surface of public discourse.
What didn't add up — what still doesn't — is the consistency problem. Creators reported wildly uneven enforcement. LGBTQ creators and fat creators documented specific patterns of suppression that seemed structural rather than random, a suspicion TikTok disputed but never fully dispelled. Creators suspected the platform's automated systems scanned audio as well as text, silently penalizing spoken words in videos even when captions were clean. TikTok never confirmed this. The uncertainty itself became a kind of pressure, forcing creators to modify not just what they typed but what they said aloud, how they spoke, what sounds they made. The algorithm's opacity meant that any word, at any time, might be the wrong one.
The Julia Fox incident crystallized the problem of collateral meaning. When coded terms spread widely enough, they stop being codes. They become words — words that carry hidden freight invisible to anyone outside the community that coined them. "Unalive" is now used by people who have no idea it was ever a workaround. "Corn" has become genuinely ambiguous in certain contexts. The euphemism "cheese pizza" — a much older coded term for child pornography, predating algospeak by decades — shows how these substitutions can calcify into permanent slang, their origins forgotten, their meanings locked in.
What investigators confirmed: algospeak is real, widespread, and structurally motivated by content moderation systems. The 2024 study finding that sentiment analysis models rated negative comments using letter-number substitution and extraneous hyphenation more positively than intended proved the evasion works — at least against older systems. The same year, a separate study found that GPT-4 could often identify and decipher algospeak, particularly when given example sentences. The arms race had a new front.
What remained contested was the platform's role in creating the conditions for all of this. TikTok maintained that enforcement was consistent and fair. Creators, researchers, and civil liberties advocates maintained it was not — that marginalized communities bore disproportionate suppression, that the system was structurally biased even if not deliberately so. No independent audit of TikTok's moderation systems has resolved this question.
What the community came to believe, and what some researchers now theorize, is that the uncertainty is the point — that keeping creators perpetually unsure of the rules forces constant linguistic adaptation, constant engagement, constant self-censorship. Whether that uncertainty is designed or simply emergent from the complexity of automated moderation at scale is a question no one has answered.
In 2025, linguist Adam Aleksic published *Algospeak*, the first book-length treatment of the phenomenon. His definition was expansive: any language change driven by digital platform constraints. Under that framing, algospeak isn't a niche behavior. It's a fundamental force reshaping how language evolves in the twenty-first century — not through poets or politicians or cultural drift, but through the silent decisions of automated systems that most users will never see or understand.
The shadow language keeps growing. New terms appear, spread, get detected, and mutate. GPT-4 can read the code now, which means the code will change. Somewhere in a comment section right now, a word means something its dictionary definition doesn't capture — a meaning shared by thousands of people who learned it from each other, in response to a machine that never explained its rules. The machine keeps watching. The language keeps moving. And the distance between what people say and what they mean keeps quietly widening.
SOURCES