The Algorithm Taught Us to Say “Unalive”

The word “unalive” used to sound like a joke. Now it appears in dictionaries. Cambridge defines the verb as a social-media substitute for “kill,” originally used so that comments or videos would not be automatically removed for violent or unsuitable content. Merriam-Webster’s slang dictionary gives the same basic history, while Dictionary.com now lists contemporary verbal and adjectival senses. “Seggs,” another platform-born workaround, has made the same journey from deliberate misspelling to dictionary entry.¹²³ That movement from evasion tactic to ordinary vocabulary is the part of algospeak that should get our attention, because it shows that language invented for a particular technological pressure can outlive the pressure that produced it.

The internet has always created slang, and every communication technology changes language in some way. Text messaging encouraged abbreviation, forums and games produced acronyms and jargon, and communities under political or social pressure have always used euphemism, metaphor and code to speak around authority. Teenagers have spent centuries inventing language adults cannot immediately decode. What is new about algospeak is the authority being negotiated with. A creator saying “unalive” may not be hiding meaning from another person at all; the intended human audience understands the word perfectly. The obscured listener is software. People are now modifying ordinary English so humans can understand a sentence while an automated classifier is, at least temporarily, less certain about it. The language is public, but one of its audiences is a machine.

The platform does not have to ban the word

The popular version of the algospeak story imagines a simple blacklist: a creator says “suicide” or “sex,” the platform hears a forbidden word, the post disappears, and users invent “unalive” or “seggs” instead. Actual moderation is more complicated. TikTok’s public rules prohibit content that promotes or instructs suicide and self-harm while allowing some prevention, recovery and supportive discussion. Its enforcement model also separates outright removal from age restriction and eligibility for recommendation in the For You feed. Sexually explicit language can violate platform rules, while other mature or suggestive material may remain online but receive restricted distribution.⁵⁶ Instagram has likewise used a layered recommendation system in which some content may stay on the service while being excluded from broad recommendations, including certain material involving self-harm, death, depression or sexual suggestiveness.⁷

For creators, those distinctions can collapse into the same experience: a post does not travel as expected. It may have been removed, age-restricted, made ineligible for recommendations, delayed for review, demonetized, shown to fewer people, or simply performed badly for reasons unrelated to moderation. The creator often cannot see which mechanism mattered. That uncertainty is central to the language change. A company does not need to maintain a literal banned-word list for users to behave as though particular words carry invisible risk.

Moderation at this scale is necessarily automated

The reason users try to speak around machines is that machines now do most of the first-pass listening. TikTok’s latest European transparency report says the platform removed roughly 104 million pieces of content between January and June 2026 across videos, live streams, ads, product listings and comments. Automated systems actioned 94.1% of violating content without human review; in the second half of 2025, TikTok reported a similar figure of 93.8%.⁸⁹ At that scale, automation is not an optional feature but the operating model.

The platform argues that automation improves speed, consistency and safety, and the basic rationale is difficult to dispute. Billions of posts cannot all wait for a person to examine them, and systems that detect graphic violence, exploitation, harassment and other prohibited material can reduce exposure to genuine harm. The linguistic problem is that automated systems classify patterns while people speak in context. The same vocabulary can appear in abuse, pornography, threat, education, reporting, recovery, satire or testimony, and a word can describe harm without causing it.

The machine hears the vocabulary before it understands the situation

That tension becomes most visible when educators, activists, journalists or survivors need to discuss the exact subjects moderation systems are designed to recognize. A 2023 study by Ella Steen, Kathryn Yurechko and Daniel Klug interviewed 19 TikTok creators about their use of algospeak. Participants described changing words not because their intended meaning violated platform rules, but because they believed automated systems were failing to distinguish context. One creator making educational material about sexual assault described altering words such as “harassment” after enforcement experiences. LGBTQ+ creators reported modifying terms such as “lesbian,” “gay,” “trans” and “queer” after perceived removals or reduced reach.¹⁰

A later study of sexual-violence disclosures found survivors using algospeak as they attempted to share experiences and build support networks inside a moderation environment where sexual language can be treated as sensitive.¹¹ The platform faces a real safety problem: malicious and benign content can share vocabulary. The creator faces a different problem: a person speaking about harm may need the exact word that describes the harm. When the system cannot reliably distinguish those uses, euphemism becomes a form of defensive writing.

The most powerful censorship may happen before moderation

If a creator knows that a particular expression will definitely violate a clearly stated rule, the boundary is at least legible. The creator can challenge it, comply with it or leave. Algorithmic moderation often operates in a murkier zone. Users do not always know what triggered a previous restriction, whether the system misunderstood a word, whether a viewer report mattered, whether an account was categorized differently, or whether poor reach had anything to do with moderation at all. Researchers call the explanations people construct around these mysteries “algorithmic folk theories”: informal models built by observing outcomes, comparing experiences and testing behavior.¹²

Theories do not have to be perfectly accurate to change speech. Studies of perceived shadowbanning among marginalized users show people collaboratively developing and testing explanations for why content appears suppressed, then changing behavior in response.¹³ A 2025 study surveying 627 UK TikTok users found marginalized users frequently felt content that did not violate guidelines was nevertheless censored or suppressed.¹⁴ A 2026 study of creators from historically disempowered groups found a recurring conclusion that understanding TikTok’s algorithms felt largely impossible, even as creators developed strategies around the patterns they believed they could see.¹⁵ The platform writes the official rulebook; users write a second one from experiments, rumors and scars.

Language becomes an arms race

Once a workaround succeeds, it begins to fail. Steen and colleagues found creators describing a cycle in which simple substitutions, such as replacing letters with numbers, could be recognized relatively quickly, while newly invented semantic substitutes seemed to work better for a time. As popular algospeak terms spread, creators assumed moderation systems learned them too, forcing another round of invention.¹⁰ Computational research supports the basic logic. A 2024 paper testing a large language model against known algospeak terms found that the system could recover the intended meaning of most terms and performed even better when given sentence context.¹⁶

The linguistic trick therefore contains its own expiration mechanism. Humans invent a euphemism because the machine does not understand it; the machine is retrained because humans are using the euphemism; humans create another one. The language changes because one of its listeners keeps learning. Algospeak is less a stable dialect than a moving border between human contextual knowledge and automated classification.

The euphemism treadmill gets a processor

Linguists have seen versions of this process before. Taboo words regularly generate euphemisms, the euphemism eventually inherits some of the discomfort attached to the original concept, and another substitute appears. Steven Pinker popularized the phrase “euphemism treadmill” for this recurring cycle, though the linguistic behavior itself is much older.¹⁷ Algospeak adds a different kind of pressure to that treadmill.

Traditional euphemism usually manages human discomfort, social convention or political prohibition. The new expression must sound safer or more acceptable to another person. Algorithmic euphemism can emerge even when nobody in the human conversation objects to the original vocabulary. “Sex” becomes “seggs” because a creator believes a corporate moderation or recommendation system may treat the original word as risky. A computational taboo has been layered onto the social taboo, and the resulting language can survive even when the original moderation tactic no longer works.

Corporate policy becomes linguistic weather

Private technology companies have always influenced speech through rules, but platform moderation combines forms of power that older gatekeepers usually kept separate. A newspaper could refuse to print a word, a television network could impose broadcast standards, and a school could prohibit profanity. Social platforms can decide what content is forbidden, what content is allowed but unsuitable for recommendation, what advertisers will support, and what ranking systems are likely to surface. A creator trying to make a living does not need to be formally banned to feel that pressure because visibility itself is economic value.

This is why the phrase “corporate censorship” is both useful and incomplete. Some moderation decisions clearly restrict expression, but much of the pressure creators describe is softer: downranking, recommendation eligibility, monetization, uncertain enforcement and the anticipation of possible consequences. The effect can still be linguistic. Opacity turns moderation into weather. Users cannot control the system, so they dress their sentences for the conditions they expect.

The coded words can outlive the code

If algospeak remained confined to creator captions, it would be an interesting platform adaptation. It has not. “Unalive” is now documented by Cambridge, Merriam-Webster and Dictionary.com as a contemporary verb or slang term, and Dictionary.com defines “seggs” as a disguise for “sex” used online to evade filters and in ordinary conversation to soften explicit wording.¹²³ Linguist and creator Adam Aleksic has argued that this vocabulary is spilling into offline speech, including school writing and ordinary conversation. In interviews around his 2025 book Algospeak, he described students using “unalive” while discussing literature and pointed to broader ways algorithmically accelerated slang moves beyond the platform that produced it.¹⁸¹⁹

Once that happens, the word no longer belongs to moderation. English does not care why a word was invented. Words enter through trade, migration, music, war, professions, advertising, jokes and mistakes; if enough people understand a term, repeat it and transmit it, it becomes part of the language regardless of origin. A temporary technological workaround can therefore leave a permanent linguistic residue.

Not every viral euphemism becomes English

The strongest claims about algospeak sometimes overreach. English is not being permanently rewritten every time a creator swaps a letter for an emoji. Most slang disappears, and many coded terms remain tightly bound to the platform culture that produced them. “Unalive” may endure; thousands of other substitutions will not. The parts of language most easily altered by this process are open vocabulary such as nouns, verbs, adjectives, slang expressions and discourse markers, while grammar is much more resistant to deliberate intervention.¹⁷

The more defensible claim is narrower but still consequential: automated moderation has become a new force in lexical evolution. It can create words, accelerate euphemisms, alter connotations and determine which expressions circulate widely enough to become familiar. The platform may not redesign English grammar, but it can exert pressure on what speakers choose to call things, especially around topics that carry safety, advertiser or recommendation sensitivity.

This is creativity under distorted incentives

“Unalive” is often mocked as evidence that internet language is becoming childish or sanitized, but that criticism mistakes adaptation for incapacity. Someone who uses algospeak usually knows the direct word; the mechanism depends on that knowledge. A creator says “seggs” precisely because the audience understands “sex.” Users exploit sound, spelling, visual resemblance, metaphor, emoji and community knowledge to preserve meaning while changing surface form. A 2026 linguistic study examining 550 instances of algospeak categorized a broad range of morphological and semantic strategies and treated them as creative forms of euphemistic and taboo-management language.²⁰

The concern is not that speakers are becoming less articulate. It is that the incentives shaping their articulation are increasingly opaque and commercial. Euphemism also changes tone. “Unalive” can create useful distance in humor or casual conversation, but in journalism, education, mental-health communication or testimony it can flatten distinctions that matter. A murder is not merely an “unaliving,” and a person who died by suicide should not automatically be described in terminology invented for platform survival. When algospeak migrates into formal settings, the pressure that created the word can disappear from view while the softened wording remains.

Bad actors speak algospeak too

Any serious critique of automated moderation has to acknowledge the mirror image of the problem. People spreading hate, exploitation, harassment or dangerous material can modify language for the same reason benign creators do: to evade detection. A moderation system that never learned euphemisms would quickly become ineffective. The 2024 ACL study showing that large language models can decode established algospeak terms was explicitly presented as a way to improve detection of moderation avoidance.¹⁶ Research published in 2026 on altered language in Chinese social-media video captions likewise found collective strategies aimed at navigating perceived moderation and censorship, illustrating how the same linguistic creativity can serve very different purposes.²¹

No responsible platform can promise never to interpret coded language. The harder demand is contextual interpretation that is accurate enough not to punish legitimate discussion while still detecting harmful evasions. That is difficult because context is computationally and institutionally expensive. A system has to distinguish a threat from a quotation of a threat, pornography from sex education, self-harm promotion from recovery discussion, hate speech from reporting on hate speech, and harassment from a survivor describing harassment. Human moderators struggle with those distinctions too. Automating them at global scale turns nuance into an engineering problem with social consequences.

The gap between rules and experience is where the dialect grows

Meta’s Oversight Board has made automation and AI a 2026 strategic priority, emphasizing both the scale of automated enforcement and the freedom-of-expression harms that inaccurate decisions can produce.²² Platform companies increasingly publish transparency reports, appeals data and policy explanations, which is a significant improvement over completely opaque enforcement. Yet detailed rules do not automatically eliminate creator fear because users experience platforms through reach, recommendations, removals and revenue rather than policy documents.

Clear rules can produce protest, but unclear systems produce superstition. If users knew with confidence that accurate discussion of suicide prevention, sexuality, race or violence would not be penalized when presented in allowed contexts, much of the incentive for defensive euphemism would weaken. Instead, creators often behave as though words themselves may carry hidden risk. The dialect develops in the gap between what the platform says its rules are and what users believe the platform has done to them.

The machine has entered the speech community

There is something historically strange about a multinational company producing a dialect without intending to. TikTok did not invent “unalive,” and Instagram did not commission “seggs.” Users built those terms collectively while adapting to systems the companies designed. The platforms created an ecological pressure, and speakers supplied the linguistic mutation. Algospeak is therefore neither purely corporate language nor purely grassroots slang. It is language formed at the border between institutional power and human improvisation.

Linguists often describe language through speech communities, groups that share norms about how expressions are used and interpreted. Platform society has added a peculiar participant. The machine does not understand language as a person does, but speakers change what they say according to what they think it understands. They anticipate classification, test phrases, monitor outcomes and teach one another what vocabulary may be safer. Automated systems have become socially meaningful listeners even when their interpretation is statistical rather than human. We speak differently because they are in the room.

This is not quite Orwellian doublespeak

The phrase “double-speak” carries an obvious political echo, but algospeak should not be forced into the frame of Orwell’s Newspeak. Social platforms generally are not ordering people to believe that killing is “unaliving” or that sex does not exist, and creators are not replacing words because a state has declared the underlying concept unspeakable. The contemporary dynamic is more decentralized, commercial and probabilistic.

What creators do instead is address two listeners at once. The human listener needs to recognize the intended meaning, while the surface form is adjusted for a classifier whose interpretation may affect distribution or enforcement. The sentence has to communicate and survive. That makes “double-speak” useful only in the most literal sense: one utterance is engineered for two very different audiences.

The lasting change is the feedback loop

Algorithms will become better at understanding coded language, and they already are, but that will not end algospeak. Humans are exceptionally good at creating context-dependent meanings because language itself is an improvisation system. A gesture, sound, joke, image or ordinary word can acquire a new meaning almost immediately if enough people share the context. Aleksic has argued that this gives human speakers a durable advantage in the cat-and-mouse game: systems can learn yesterday’s euphemism while communities are already building tomorrow’s.¹⁸

Whether “unalive” remains ordinary English 30 years from now is almost beside the point. The deeper change is already embedded in the communications environment. Enormous populations now speak inside systems where private automated tools continuously classify, rank and monetize language, while speakers continuously adapt in response. The platform detects language, users change language, the platform learns the change, and some of the new vocabulary spreads to people who were never trying to evade anything. Eventually dictionaries record the residue.

The language remembers the pressure

Every age leaves its institutions inside its vocabulary. We retain words shaped by printing, television, bureaucracy, advertising and computers long after the technologies or policies that produced them have faded into the background. Algospeak may become another archaeological layer. Future speakers could use “unalive” because it feels funny, indirect or ordinary, without knowing that millions of people once adopted the expression because they were uncertain how automated moderation would treat direct language about death.

The remarkable thing is not that corporations have finally learned how to control English; they have not, and English is far too unruly for that. The more interesting fact is that automated corporate systems have become powerful enough to give millions of people the same reason to invent around them at the same time. The algorithm did not write the new vocabulary. It created the pressure, and humans did what language has always allowed them to do: they found another way to say it.

Reporting & references

1. Cambridge Dictionary, “unalive,” current definition and usage history. Source
2. Merriam-Webster Slang, “unalive.”
3. Dictionary.com, “seggs” and contemporary “unalive” entries. Source
4. Collins English Dictionary / Webster’s New World, “algospeak,” current definition.
5. TikTok Community Guidelines, Mental and Behavioral Health: Suicide and Self-Harm. Source
6. TikTok Community Guidelines, Sensitive and Mature Themes / recommendation eligibility.
7. Instagram Help Center, Recommendations on Instagram: categories allowed but potentially ineligible for recommendation. Source
8. TikTok, seventh EU Digital Services Act transparency report, Aug. 31, 2026.
9. TikTok, sixth EU Digital Services Act transparency report, April 29, 2026.
10. Ella Steen, Kathryn Yurechko & Daniel Klug, “You Can (Not) Say What You Want,” Social Media + Society, 2023.
11. “A Critical Discourse Analysis of Sexual Violence Survivors and Censorship on the Social Media Platform TikTok,” Journal of Interpersonal Violence, 2024.
12. Owen Alinsangan Doyle, “Algorithms and the ‘Anti-Preference’: A Quantitative Investigation of Reaching the Wrong Audience on TikTok,” 2024.
13. Daniel Delmonaco et al., “‘What are you doing, TikTok?’: How Marginalized Social Media Users Perceive, Theorize, and ‘Prove’ Shadowbanning,” PACM HCI, 2024.
14. Eddie L. Ungless, Nina Markl & Björn Ross, “Experiences of Censorship on TikTok Across Marginalised Identities,” ICWSM, 2025.
15. Sarah Florini et al., “Algorithmic folk theories: Content creators’ understandings of TikTok’s algorithms,” First Monday, 2026.
16. Jan Fillies & Adrian Paschke, “Simple LLM based Approach to Counter Algospeak,” ACL Workshop on Online Abuse and Harms, 2024.
17. Steven Pinker, discussion of the “euphemism treadmill” and linguistic change, Conversations with Tyler.
18. Reuters, “Adam Aleksic on how algorithms are transforming the way we communicate,” July 24, 2025.
19. NPR / Weekend Edition, Adam Aleksic interview on Algospeak and social-media-driven language change, July 20, 2025.
20. “Censorship in the age of social media and algorithms: A linguistic study of TikTok algospeak,” Studia Neophilologica, 2026.
21. Andy Zhao, “Altering Words to Evade Perceived Moderation: Decoding Algospeak in Chinese Social Media Video Captions,” ICWSM, 2026.
22. Meta Oversight Board, 2026 strategic priority on Automation and AI in content moderation.