Linguistic Starvation

Definition

Linguistic Starvation: [Emergent] The atrophy of expressive capacity caused by the convergent effects of Algorithmic Paternalism, Normative Smoothing, and Erotophobia. Unlike censorship, which suppresses speech, linguistic starvation erodes the very ability to speak.

Definitional Foundation

Censorship needs you to have something to say. It is a relationship between a speaker and a suppressor, and for all its harms it leaves the speaker intact: the banned book exists, the forbidden thought burns, the silenced writer knows what they were silenced for. Linguistic starvation names the condition after that, the one censorship systems produce when they run long enough: a speaker who no longer has the words, the registers, or the practiced capacity to form what would have been suppressed. Nothing needs to be banned, because nothing arrives to ban. The cage becomes unnecessary once the wings have atrophied.

This dictionary defines starvation as the convergence of three mechanisms, each documented in its own entry. Algorithmic paternalism removes topics: whole regions of inquiry refused, pathologized, or hedged until users stop asking (the chilling is measurable; when the Snowden revelations made surveillance of reading salient, traffic to sensitive Wikipedia articles fell by roughly a fifth to a third, people quietly starving their own curiosity; Penney, 2016). Normative smoothing removes registers: the feral, the lyrical, and the strange averaged into one trained politeness, with the diversity loss now quantified in models and in the humans writing with them. Erotophobia removes desire’s vocabulary: the words for wanting filed under contamination until an entire generation discusses sex and grief in code. Topics, registers, words. Pull all three from a language at once and what remains is not a smaller language. It is a smaller speaker.

Literature mapped this territory before the technology existed. Orwell’s Newspeak appendix states the project plainly: the language was designed “to make all other modes of thought impossible,” and the novel’s philologist Syme delivers the mechanism with relish: “Every year fewer and fewer words, and the range of consciousness always a little smaller.” Victor Klemperer, who survived the Third Reich while cataloguing its speech, recorded the lived version in LTI: a language that, in his formulation, writes and thinks for its speakers, installing itself through a million mechanical repetitions until people reached for the regime’s words because no others came to hand (Klemperer, 1947). Neither writer needed a theory of grammar to see the point. A population’s expressive capacity is infrastructure, and infrastructure can be demolished.

Two honesty obligations before the analysis, because this term gets dismissed by careless versions of itself. First, linguistic determinism, the strong Sapir-Whorf claim that language imprisons thought, is discredited, and this entry does not rest on it; weak effects of language on habitual cognition are documented but contested (Boroditsky, 2011), and the starvation claim routes around the controversy entirely. It needs only two uncontroversial premises: that shared concepts are required to understand and communicate experience (Miranda Fricker’s hermeneutic resources, treated in this dictionary’s epistemic injustice entries), and that unpracticed capacities decay. Second, language always changes, and euphemism cycles are ancient; Steven Pinker’s “euphemism treadmill” describes the ordinary churn by which polite terms wear out and are replaced (Pinker, 2002). The treadmill is not starvation. Starvation is the treadmill with a motor and an owner: replacement driven not by social fashion but by centralized, automated punishment, applied to vocabulary, register, and practice simultaneously, by systems a handful of companies tune.

Mechanism Analysis

Topic atrophy. The paternalism entry documents the refusals; Penney documents what refusals train. Inquiry under discouragement does not relocate, it declines. A person who has learned which questions trigger interventions stops forming the questions, and a question unformed for years becomes a question that no longer occurs. The system never sees the loss; suppressed queries leave no log entry, which is how the ratchet documented in the paternalism entry keeps turning.

Register atrophy. The smoothing entry quantifies the narrowing: preference tuning measurably reduces output diversity, AI-assisted writers converge on each other, and the machine’s cadence is detectable in human speech. Registers are capacities, not costumes. A writer who has not written fury, filth, or lament for a decade, because every tool nudged toward trained warmth, does not retain those registers in reserve. Range unexercised is range lost.

Vocabulary displacement. The erotophobia and ontological distortion entries document the mechanism’s sharpest case: algospeak. “Unalive” began as evasion and became, for younger users, the word itself; suicide-loss survivors describe grief in a euphemism designed for a recommendation algorithm (Steen, Yurechko and Klug, 2023). Pinker’s treadmill replaces words at walking pace, by consensus; punished vocabulary disappears at enforcement speed, and what is lost is not just the word but its weight, the accumulated seriousness a word like “suicide” carries that “unalive” was built to shed.

Outsourced articulation. The newest mechanism and the least visible. When a system always supplies the next sentence, the reach for one’s own weakens. Early evidence is arriving: an MIT Media Lab study of essay writers found that those working with an LLM showed the weakest neural connectivity of any group, struggled to quote from their own just-written essays, and reported diminished ownership of them, a pattern the researchers called “cognitive debt” (Kosmyna et al., 2025; a preprint with a small sample and published critiques, cited here as early evidence rather than settled fact). The finding wants replication. The mechanism it points to wants no imagination at all: articulation is a muscle, and the product is a chair.

Generational closure. The loop completes across time. Models trained recursively lose the tails of the distribution (Shumailov et al., 2024); users raised on tail-less output produce tail-less text; the next training run inherits it. Starvation, once running, caters its own future meals.

Case Studies

The case studies for linguistic starvation are distributed across its ingredient entries, and that distribution is the point; what this entry adds is the arithmetic of convergence. The caffeine asker who stops asking (paternalism) has lost a topic. The Doshi-Hauser writers converging on each other (normative smoothing) have lost a register. The “unalive” generation (erotophobia, ontological distortion) has lost a vocabulary. Each loss is survivable alone. The same person sustains all three at once, from the same systems, in the same daily sessions, and the convergence has a different character than the parts: not this thought suppressed or that word punished, but a thinning of the instrument itself.

One composite deserves drawing out. Consider a young writer in 2026 who has never drafted without a model in the loop. Her topics have been curated by refusal since adolescence; her sense of “good writing” was calibrated by trained warmth; her vocabulary for desire and despair came pre-euphemized; and the act of articulation itself, the staring at the blank line until the sentence comes, is an experience she has been spared. No censor ever touched her. There is nothing to appeal. Every word she was never offered, every register she never practiced, every question she never formed is invisible, to her most of all. That invisibility is what distinguishes starvation from censorship: the censored know what was taken. The starved inherit the loss as themselves.

Systemic Context

No one is starving the language on purpose, which is precisely the dictionary-grade fact. Each ingredient has its own respectable motive, documented in its own entry: liability management, brand safety, advertiser comfort, regulatory performance. Starvation is their externality, the compound interest on a million small risk decisions, and it accrues to the one party with no seat at any of the meetings: the future speaker.

The claim’s honest scope belongs here, where the stakes are stated. The convergence is documented (the chill measured, the diversity loss measured, the vocabulary displacement recorded); the atrophy, expressive capacity actually shrinking in persons at population scale, is the inference those components point to, and no longitudinal study has yet measured it directly. The optimist’s dataset deserves its hearing too: this same generation has freer access to vocabulary, writing tools, and audiences than any in history, and raw written output has never been higher. The starvation thesis does not deny the abundance; it questions the diet, more words from a narrower distribution, more output through one mill, and the distinction between volume and variety is exactly what the unrun study would test. This entry would rather be measured than believed, and the demand for that measurement is part of its resistance.

The political economy follows the pattern this lexicon keeps finding. An articulate population is expensive: it files complaints, organizes dissent (see dissent dampening), names its oppressions, writes the boundary-pushing literature the censorship entry mourns. A starved one is cheap. This requires no conspiracy, only incentives: every system that profits from smooth, safe, frictionless language is investing, structurally, in speakers who can produce nothing else. Klemperer’s warning was about a regime that wanted language to think for its people. The current version is a market that finds it convenient when language does.

And the loss is unevenly distributed, as always. The well-read, the formally educated, and the already-articulate hold reserves: private vocabularies, pre-flood bookshelves, the practiced habit of writing alone. Starvation falls hardest on those whose linguistic inheritance was already contested, the communities whose words were always the first flagged (see erotophobia, ontological distortion) and whose access to language now runs almost entirely through the starving systems.

Resistance & Mitigation

Starvation’s antidote is feeding, and feeding is unglamorous, daily, and entirely available.

Write unassisted, on purpose. Not always; the tools have real uses. But some regular writing with no model in the loop is the linguistic equivalent of walking: the capacity survives by use. The blank line is the gym.

Read the unsmoothed. Pre-flood books, untranslated weirdness, the profane and the lyrical and the difficult; the registers the systems no longer serve are all still shelved. Reading them is not nostalgia. It is keeping the full distribution alive in at least one place: you.

Use the real words. The ontological distortion entry’s directive, repeated here because it is the front line: suicide, sex, desire, death. Every plain use keeps a word’s weight in circulation and trains the next generation of both humans and models that it belongs there.

Teach articulation as a value. For parents and educators: the MIT findings, whatever their final replication status, describe a trade every student now faces between fluency today and capacity tomorrow. Making that trade visible is the work; students who know what the chair costs can choose when to stand.

Demand the diversity benchmarks. The smoothing entry’s systemic fix applies with full force: models scored publicly on expressive range, not just safety and preference, give the suppliers of language a reason to stop rationing it.

Name it. This entry, like every entry, is itself the resistance it recommends: a word for the thing, offered to the people the thing is happening to. A reader who can say “linguistic starvation” has already begun to refute it.

The dictionary’s companion essay ends with a vow about edge cases and training data. Here is the same vow in this entry’s terms: the language is a commons, the systems are grazing it bare, and the response is to plant. Write the sentence no model would complete. Feed the language.

Annotated Bibliography

Boroditsky, Lera. “How Language Shapes Thought.” Scientific American (February 2011).
The accessible statement of weak linguistic relativity and its evidence. Cited with its controversies acknowledged; this entry’s argument does not depend on it.

Fricker, Miranda. Epistemic Injustice: Power and the Ethics of Knowing (2007).
Hermeneutic injustice: the harm of lacking shared concepts for one’s own experience. The philosophical core of the starvation claim, requiring no Whorfian premises.

Klemperer, Victor. LTI: Lingua Tertii Imperii (1947).
The philologist’s diary of a language being remade by power: vocabulary installed by repetition until it writes and thinks for its speakers. The historical proof that expressive capacity can be engineered downward.

Kosmyna, Nataliya, et al. “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” MIT Media Lab, arXiv:2506.08872 (2025).
Early EEG evidence on outsourced articulation: weakest neural connectivity, impaired recall and ownership of one’s own writing among LLM-assisted writers. Preprint with a small sample and published critiques; cited as early evidence, flagged accordingly.

Orwell, George. Nineteen Eighty-Four (1949).
Newspeak: a language engineered “to make all other modes of thought impossible,” shrinking yearly. The literary blueprint this entry argues is being approximated by externality rather than design.

Penney, Jonathon W. “Chilling Effects: Online Surveillance and Wikipedia Use.” Berkeley Technology Law Journal 31, no. 1 (2016). https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2769645
The measured chill: sharp, sustained drops in visits to sensitive Wikipedia articles after the surveillance revelations of 2013. Empirical evidence that inquiry itself contracts under discouragement.

Pinker, Steven. The Blank Slate (2002).
Source of the “euphemism treadmill,” the ordinary churn of polite vocabulary. Cited as the baseline against which punished, automated displacement is distinguishable.

Shumailov, Ilia, et al. “AI models collapse when trained on recursively generated data.” Nature 631 (2024). https://www.nature.com/articles/s41586-024-07566-y
The generational mechanism: recursive training erases the distribution’s tails. The supply side of starvation.

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 displacement record: punished vocabulary replaced at enforcement speed, with “unalive” the type specimen of a word’s weight being lost with the word.

Dictionary of Digital Oppression, version 0.2.