Maybe you’ve heard it happen. Someone is telling a story at dinner, and a suicide enters the story, and the word they reach for is “unalive.” No phone is open. No caption is being scanned. The content filter that taught them the word is nowhere in the room, and they defer to it anyway.
“Unalive” belongs to a dialect now known as algospeak — the substitute vocabulary platform users invent to keep automated moderation systems from suppressing, restricting or deleting what they post. When researchers Ella Steen, Kathryn Yurechko and Daniel Klug interviewed nineteen TikTok creators for a 2023 study published in Social Media + Society, they came back with a working lexicon. The corn emoji stands in for porn. “Seggs” stands in for sex, having replaced “s3x” once the leetspeak version seemed to stop working. “Le$bean” — say it aloud and it becomes “le dollar bean,” a small joke folded into the disguise — stands in for lesbian. In the weeks after the Supreme Court overturned Roe v. Wade in 2022, abortion became “shmortion.” Even the platform itself gets a code name, “the clock app,” as if naming TikTok might summon its attention.
The list nobody has seen
Evasive slang is an old technology. Pickpockets had cant; the French banlieues built verlan. But every earlier version was developed against a human authority whose habits could at least be studied, and the striking thing about algospeak is that its authority is invisible — and possibly, in places, imaginary. The creators in the TikTok study were not reciting rules. They were guessing them. One described a literal “list of unapproved words” and warned that if you say the word “sex,” you will get banned; the researchers are careful to frame such claims as folklore, inferred from moderation users experience as non-contextual, random and biased against marginalized communities — not as published policy anyone could consult.
When the rules are secret, superstition does the work of law. Users become reverse engineers of a machine they cannot see, swapping rituals the way gamblers swap lucky socks. One creator estimated her evasions worked about ninety percent of the time — a self-diagnosis rather than a measurement, but it captures the mental arithmetic: the ritual costs almost nothing, and the imagined penalty is a dead video, a throttled account, a vanished income. It is entirely possible that some of the rules being so carefully dodged do not exist. The dialect is no smaller for it. Knock on wood.
What is genuinely new, beyond the opacity, is the pace and the audience. The euphemism treadmill — the cycle by which polite substitutes absorb the stain of whatever they replace — used to run on a generational clock. On TikTok it runs in weeks: s3x worked until, users believed, the filter learned to read it, and seggs reported for duty. And every previous dialect in history evolved to be understood by some people and missed by other people. Algospeak is likely the first engineered to be misread by a machine — instantly legible to your followers, harmless-looking to the classifier standing between you and them. Every caption now has two readers, and only one of them forgives.
The words that need disguises
It is worth pausing on which words get the treatment. Not “crypto,” not “brunch.” The documented algospeak lexicon clusters, with grim consistency, around the heaviest territory in the language: sex, pornography, abortion, suicide and homicide (both filed under “unalive”), sexual assault (“mascara,” according to Associated Press reporting, a substitution that files survivors’ testimony under cosmetics), and identities like lesbian and autistic (“awetistic”) — communities the research suggests already face biased moderation. The subjects for which plain language matters most, where someone is searching for help or testifying or teaching, are precisely the subjects being euphemised.
The costs are not abstract. Search runs on plain words, and a teenager trying to learn something real about her own body will struggle to find the conversation if the conversation has moved into code. In the TikTok study, one creator had a video’s audio deleted for naming a sex toy — a sex-education casualty of a system that cannot reliably tell teaching from soliciting. A second study of algospeak on TikTok found that the evasions helped users avoid moderation consequences but often impeded the creation of quality content; you escape the filter by making the thing slightly worse. And then there is the tonal cost nobody chose: these words are goofy. They arrive with a wink built in, and the wink gets imported into suicide, rape and grief. “Unalive” makes a death sound like a plot point. Andrea Beltrama, a linguistics researcher at the University of Pennsylvania, has called algospeak “lexical innovation,” which it plainly is — but innovation describes the mechanism, not the outcome, and nobody voted on this particular update.
None of this would matter much if the dialect stayed where it was born. But vocabulary is not a uniform you change out of at five o’clock; you do not keep one lexicon for the app and another for the rest of your life. The words you type all day become the words that come out of your mouth, which is how a term coined to sneak past a moderation system ends up spoken at a dinner table where no moderation system is present. Self-censorship, once trained, does not clock out. The filter never has to follow anyone home — it is already there, working in the first person, free of charge.
Perhaps the strangest part of the arrangement is that the list of unapproved words has never been produced. It may not exist in any form a person could read. Yet it is being written anyway, in negative space, by everyone it governs — every corn emoji, every shmortion, every “unalive” spoken aloud in a room with no algorithm in it is another entry in a dictionary compiled entirely by avoidance. The platforms never published their forbidden vocabulary. They didn’t have to. We are keeping the list for them, and increasingly we speak from memory.