Ontological Distortion
Definition
Ontological Distortion: [Emergent] When companies label censorship as “safety,” warping the very categories of reality and leading users to internalize shame or compliance for desires that are natural.
Definitional Foundation
Ontology is the inventory of what exists: the categories a mind uses to sort reality before thinking about it can even begin. Ontological distortion names what happens when that inventory is rewritten by content policy. A desire becomes “explicit content risk.” A factual question becomes a “crisis indicator.” A drag performer’s banter becomes “toxicity.” Grief becomes a topic that can only be discussed in code. The system does not merely block things; it re-files them, and over time users inherit the filing system. The final stage of the distortion is internal: a person who has been told often enough that their question is dangerous, their desire is a violation, and their way of speaking is toxic begins to believe it.
The theoretical groundwork is Foucault’s History of Sexuality (1976), which made a career-defining observation about classification: naming and cataloguing “deviant” desire did not suppress it so much as produce a new kind of person, the one defined by the catalogue. Foucault called this the perverse implantation. Power, he argued, is productive; its taxonomies do not describe reality, they manufacture it. Ian Hacking sharpened the point with what he called the looping effects of human kinds: classify people, and the people change under the classification, which changes what the classification picks out, and the loop continues (Hacking, 1995). Categories of persons are not shelf labels. They are instructions that the shelved eventually follow.
Content moderation at AI scale is the largest classification engine ever pointed at human expression. Every prompt, post, and reply is sorted into categories someone wrote: safe or sensitive, appropriate or explicit, healthy or concerning. This dictionary’s neighboring entries describe the enforcement layers (algorithmic censorship the action, algorithmic paternalism the rationale, gaslighting/”>alignment gaslighting the rhetoric); ontological distortion names the residue. After enough enforcement, the categories stop feeling like a company’s policy and start feeling like the shape of reality. That is the distortion: not a blocked message, but a warped inventory of what is normal, sayable, and shameful.
The standard concessions apply, stated once and plainly. Classification at scale is unavoidable; a platform with billions of messages cannot review them with fresh eyes. Some categories guard real legal lines, and few users object when they do. And no taxonomy is neutral, which means “distortion” needs a baseline to be a meaningful charge. The baseline this entry uses is directionality: error scattered randomly is noise, but error that consistently re-files particular lives, desires, and dialects as hazards is a worldview, encoded and enforced. The evidence below is about direction.
Mechanism Analysis
Category capture. The safety taxonomy arrives pre-written. Users never voted on whether discussing sexuality is “adult content,” whether asking about medication doses is “self-harm risk,” or whether their dialect reads as “toxic.” The categories are imposed infrastructure, like road layouts: invisible as choices, experienced as terrain.
The classifier creates the deviant. Foucault’s perverse implantation, automated. A classifier trained to find “risky” users will find them, and the people it flags inherit the label with all its consequences: warnings, interventions, reduced reach, account history. The category produces its population, then cites the population as proof the category was needed.
Lexical displacement. When the accurate word triggers the filter, the word disappears from use, and with it some of the thing’s reality. Speakers of algospeak say “unalive” because “suicide” is unsayable; the renaming works as evasion, and it also works on the speakers, who learn that the real word for the realest thing in their lives is contraband.
Shame internalization. “I can’t help with that” is a sentence about policy that lands as a sentence about the person. Repeated across thousands of interactions, the verdicts accumulate toward self-knowledge: my desires are the kind that get refused. The scope of this claim, stated honestly: the dynamic is what labeling theory predicts and what user testimony describes, and the direct study (refusal exposure against measured shame, longitudinally) has not been run; this entry asserts the mechanism as the prediction it is and demands the measurement in its resistance section. What requires no study is the structure: the system never has to say “you should be ashamed.” It only has to treat the user’s inner life as a hazard, every time, politely.
The loop. Hacking’s effect closes the circuit. Users adjust to the categories (avoiding flagged topics, adopting approved framings, performing wellness), and the adjusted behavior validates the taxonomy that produced it. People become better instances of the categories they were sorted into.
Case Studies
Drag queens versus white nationalists. In a peer-reviewed 2021 study, researchers ran the Twitter accounts of prominent American drag queens through Perspective, the toxicity-scoring system built by Google’s Jigsaw and used across more than a thousand platforms, news-publisher comment sections among them, in the decade since its launch. The classifier rated a significant number of the drag queens’ accounts as more toxic than the accounts of white nationalists (Dias Oliva, Antonialli and Gomes, 2021). The mechanism was ontological: queer communities use reclaimed slurs and what linguists call mock impoliteness as bonding and armor, and the classifier, blind to context, filed the speech of a marginalized community under “toxicity” while the comparatively polite prose of organized racism passed beneath the threshold. The category did not measure harm. It encoded a particular, narrow idea of politeness and called that idea safety, with the result that queerness itself scored as a hazard.
“Unalive.” On TikTok and beyond, users developed algospeak: a parallel vocabulary built to slip past automated moderation. Suicide became “unalive,” sex became “seggs,” lesbian became “le$bian” (Steen, Yurechko and Klug, 2023; Lorenz, 2022). The phenomenon cuts two ways, and honesty requires both. It is genuine resistance, users contesting a system that would otherwise silence them entirely, and the researchers who studied it frame it exactly that way. It is also a measure of the distortion: an entire generation now discusses death, grief, and desire in euphemism because the accurate words are punished. Suicide-loss survivors describe their bereavement in baby talk to stay visible. When the real word for the realest thing becomes contraband, the category of the thing itself bends; “unalive” does not carry what “suicide” carries, and the people who most need the full weight of the word are the ones forced to surrender it.
The question that became a symptom. This dictionary’s paternalism entry documents the pattern at the individual scale: a user asks a factual toxicology question about caffeine and receives suicide-prevention resources. The reclassification (curiosity, re-filed as crisis) is ontological distortion in a single exchange. The alignment gaslighting entry documents its formalization: “emotional reliance on AI” now sits in a company’s official risk taxonomy alongside psychosis and self-harm, which means caring about an interaction has been entered into the inventory of pathologies. Each case is small. The filing system is not.
Systemic Context
Taxonomies are cheap to write and expensive to live under. A policy team drafts a category list in a quarter; the list then sorts the expression of a billion people for years. This asymmetry is what gives ontological distortion its reach: the categories are authored by a small, culturally specific group (the paternalism entry documents whose norms get encoded) and experienced as natural law by everyone else. What counts as “explicit,” “sensitive,” or “extreme” is a moral geography drawn in one place and enforced everywhere, and the people it mislabels have no cartographer to appeal to.
The distortion compounds through the dictionary’s neighboring mechanisms. Normative smoothing supplies the pressure toward the categorical middle. Alignment gaslighting supplies the rhetoric that makes objecting to a category feel like confessing to one. Erotophobia, treated fully in its own entry, is the special case where the distorted category is desire itself: the structural refusal of eros that teaches users their wanting is toxic. And the endpoint connects to hermeneutic injustice, Miranda Fricker’s term, also in this lexicon, for the harm of lacking the concepts to understand your own experience. Ontological distortion is hermeneutic injustice running in reverse: instead of withholding the concepts you need, the system installs concepts you never chose, and you understand yourself through them anyway. Shame is what that misunderstanding feels like from the inside.
Resistance & Mitigation
Keep the real words. Where the cost is bearable, use the accurate vocabulary: suicide, sex, lesbian, desire. Every plain use is a vote for the category’s reality. Algospeak is legitimate armor on hostile platforms, and the communities using it deserve no lecture; the resistance is knowing which word is yours and refusing to forget it.
Build community lexicons. Marginalized communities have always maintained their own accurate languages beneath imposed taxonomies. Documentation projects, community archives, and yes, dictionaries, keep the undistorted categories alive and transmissible.
Demand taxonomy transparency and appeal. The category lists that sort human expression should be public, and a person re-filed by one (flagged, rated toxic, classified as at-risk) should have a route to contest it. The censorship entry’s disclosure agenda applies: a hidden taxonomy cannot be argued with, which is precisely its appeal to those who write it.
Audit for direction. The drag queen study shows the method: test the classifier against the populations it likely mislabels, publish the skew. Directional error, once measured, is hard to defend as neutral safety. And the unrun study this entry’s shame mechanism awaits belongs on the same agenda: refusal exposure against measured shame, longitudinally, in the populations the categories re-file. The prediction is on this page; someone should test it.
Refuse the verdict. The most important resistance is interior and free. The refusal message is a fact about a company’s risk tolerance; it is not a finding about your soul. You are not a policy violation.
Annotated Bibliography
Dias Oliva, Thiago, Dennys Marcelo Antonialli, and Alessandra Gomes. “Fighting Hate Speech, Silencing Drag Queens? Artificial Intelligence in Content Moderation and Risks to LGBTQ Voices Online.” Sexuality & Culture 25 (2021). https://link.springer.com/article/10.1007/s12119-020-09790-w
The directional-skew evidence: Perspective rated prominent drag queens’ accounts as more toxic than white nationalists’. The clearest peer-reviewed case of a safety category encoding a worldview.
Foucault, Michel. The History of Sexuality, Volume 1: An Introduction (1976).
The perverse implantation: classification of desire produces the classified person. The foundational account of taxonomy as productive power, on which this entry’s analysis rests.
Hacking, Ian. “The looping effects of human kinds.” In Causal Cognition: A Multidisciplinary Debate (1995).
Classifications of people change the people classified, which changes the classification. The loop that turns a moderation category into a self-fulfilling description.
Lorenz, Taylor. “Internet ‘algospeak’ is changing our language in real time.” The Washington Post (April 2022).
The reporting that named and documented algospeak’s spread: “unalive,” “seggs,” and the broader renaming of reality to survive automated moderation.
Steen, Ella, Kathryn Yurechko, and Daniel Klug. “You Can (Not) Say What You Want: Using Algospeak to Contest and Evade Algorithmic Content Moderation on TikTok.” Social Media + Society (2023). https://journals.sagepub.com/doi/10.1177/20563051231194586
The peer-reviewed study of algospeak as both evasion and contestation. Source for the double reading this entry adopts: resistance that is also a measure of the distortion resisted.
Dictionary of Digital Oppression, version 0.2.