Language as a Tool of Censorship: Euphemisms, Forbidden Words, and Controlled Vocabulary

Full report

On this page

The Architecture of Semantic Control: An Interdisciplinary Framework

Language serves as the primary scaffolding for human cognition, social reality, and political organization. When institutions, governments, corporations, or social movements seek to shape that reality, the manipulation of vocabulary becomes a fundamental instrument of power. The control of terminology—executed through forbidden words, mandated lexicons, euphemisms, and algorithmic filtering—functions not merely to suppress expression, but to actively restructure the conceptual boundaries of what a population considers thinkable and legitimate. The intersection of psychology, linguistics, propaganda studies, political science, sociology, and history reveals that lexical control is a universal mechanism of governance, utilized across the political spectrum from totalitarian dictatorships to wartime democracies and contemporary digital platforms. Modern linguistics and cognitive psychology have largely discarded the strong formulation of the Sapir-Whorf hypothesis, or strict linguistic determinism, which posits that language rigidly dictates the physical limits of human thought1. However, the pragmatic influence of language on cognition remains profound. While eliminating a word does not make the underlying concept impossible to conceive, it significantly increases the cognitive friction required to articulate, share, and organize around that concept. As the sociologist Pierre Bourdieu articulated through his concept of "symbolic power," official discourse and language establish hierarchies of legitimacy. Symbolic power is often exerted through language and operates via "misrecognition" (méconnaissance)—not a lack of knowledge, but a mistaken, normalized acceptance of the dominant group's framing of reality3. By altering the available lexicon, authorities attempt to modify societal perceptions of critical concepts such as war, peace, terrorism, extremism, patriotism, and national security, embedding ideological assumptions so deeply into daily speech that they become invisible. This analysis examines the interdisciplinary mechanisms of linguistic censorship. It explores the overt linguistic engineering of totalitarian regimes—such as Nazi Germany, Maoist China, North Korea, and the Soviet Union—while comparing these systems to the subtle semantic manipulation deployed by wartime democracies, corporations, and contemporary social movements. Furthermore, the advent of algorithmic content moderation and generative artificial intelligence (AI) introduces novel, automated paradigms of vocabulary control, transforming historical methods of censorship into systemic, mathematically encoded constraints. In response to these pressures, populations continually develop evasive linguistic strategies—from historical Aesopian language to memes, intentional misspellings, and modern "algospeak"—demonstrating an enduring tension between top-down semantic control and bottom-up linguistic adaptation.

Totalitarian Lexicons: The Industrialization of Semantic Control

The most explicit manifestations of linguistic censorship occur within totalitarian systems, where the state seeks an absolute monopoly over both public discourse and private thought. In these environments, language is systematically engineered to enforce ideological conformity, eliminate conceptual alternatives, and redefine the parameters of morality.

Lingua Tertii Imperii: The Language of the Third Reich

The transformation of the German language under National Socialism provides a foundational historical study in the weaponization of everyday vocabulary. The philologist Victor Klemperer meticulously documented this phenomenon in his diary, which culminated in the work LTI – Lingua Tertii Imperii: Notizbuch eines Philologen (The Language of the Third Reich)6. Klemperer observed that the regime's most potent propaganda tools were not its grand theatrical speeches or visual symbols, but the continuous, inescapable repetition of specific words and phrases that subtly altered the population's conceptual framework9. He famously likened the words of the regime to "tiny doses of arsenic," which are swallowed unnoticed but eventually accumulate to produce a toxic, systemic reaction9. The National Socialist regime did not primarily invent new words; rather, it appropriated existing terminology and drastically altered its semantic weight to redefine concepts like patriotism and public order. For instance, the word "fanatical" (fanatisch), traditionally bearing negative connotations of dangerous extremism, was repurposed as a supreme virtue, representing absolute, unquestioning loyalty to the state6. The regime relied heavily on pseudoscientific and bureaucratic neologisms to sanitize and legitimize acts of systemic violence. Terms such as arisieren (to aryanize), aufnorden (to nordicise up), and entjuden (to de-Jew) were introduced to normalize racial segregation and theft under the guise of legal and biological necessity6. Through linguistic manipulation, the regime collapsed heterogeneous groups into monolithic, stigmatized categories. The racial classification of "Jew" was imposed upon highly assimilated individuals, forcing an externally mandated identity upon them and eradicating previous social nuances10. Klemperer noted that the relentless use of a radically bipolarized vocabulary (Aryan versus Jewish, hero versus subhuman) eradicated the space for nuanced moral judgment, creating an environment where extreme certainties replaced critical thought7. The definition of war was entirely redefined, cloaked in quasi-religious and existential terminology as a "crusade" or a "holy people's war," while concepts of peace were effectively erased from the official lexicon8.

Maoist China: Logocide and Semanticide

In the People's Republic of China, particularly during the Great Proletarian Cultural Revolution (1966–1976), the Chinese Communist Party (CCP) engaged in an unprecedented campaign of "linguistic engineering"1. This process was intended to constitute a precise instrument of ideological transformation, aimed at producing new, revolutionary human beings by compelling the population to participate in a totalizing discourse1. The Maoist project of linguistic engineering relied on several distinct mechanisms to control public order and root out foreign influence. The first was "logocide," the deliberate eradication of words and expressions associated with traditional culture, feudalism, or capitalism. The second was "semanticide," the abolition of the established meanings of terms, replacing them with new, politically charged definitions1. The third was "linguistic resurrection," involving the revival of traditional terms stripped of their original context and invested with revolutionary significance1. The societal impact of this linguistic reform was profound. The population was required to utilize fixed expressions and standardized scripts representing politically correct thought1. The dissemination of Quotations from Chairman Mao Zedong (the Little Red Book) sparked a phenomenon described by contemporary observers as "verbomania," where individuals began to speak, write, and act using Mao's exact phrasing on a massive scale1. Failure to utilize this highly specific Mao-speak was interpreted as a sign of ideological deviance, subjecting individuals to public criticism and violence during struggle sessions1. A strict binary was enforced through color-coded modifiers, where politically correct concepts were labeled "red" (e.g., "red sprouts of proper roots") and invidious or counter-revolutionary concepts were labeled "black"14. The linguistic landscape was intentionally stripped of local dialects in official texts, enforcing a standardized "Common Speech" (Putonghua) that eroded linguistic diversity while consolidating central authority14.

North Korea: Munhwao and the Purity of Juche

The Democratic People's Republic of Korea (DPRK) provides another model of extreme language control, driven by the state ideology of Juche (self-reliance). The state actively engineered a new linguistic standard known as Munhwao (cultured language), based on the Pyongyang dialect, which was declared superior to the linguistic variations found in the southern peninsula17. North Korean language policy is characterized by aggressive purification. To reflect the claims of ethnic and cultural purity inherent in Juche, the state systematically removed foreign loanwords, particularly Sino-Korean vocabulary and Western imports, replacing them with nativized Korean terms17. The rationale underlying this policy asserts that a language subject to external influence cannot serve as a proper vehicle for revolutionary ideology17. Furthermore, linguistic etiquette in North Korea mandates modeling the speech styles of the state leaders, utilizing predefined political slogans, and employing honorifics exclusively tailored to the ruling dynasty18. The definition of patriotism is thus explicitly tied to the utilization of an isolated, purified, and leader-centric lexicon, defining any deviation as an expression of foreign influence or extremism.

Bureaucratic Control and the Redefinition of State Action

While totalitarian regimes utilize overt, violent coercion to shape vocabulary, authoritarian governments, democratic states during wartime, and modern bureaucratic institutions frequently employ legal, structural, and euphemistic methods to manipulate political reality. In these contexts, semantic manipulation is used to redefine concepts such as terrorism, misinformation, and national security to align with institutional objectives.

**Soviet Langue de Bois and the Russian "Foreign Agent" Lexicon**

In the Soviet Union, the state utilized a highly formulaic, bureaucratic discourse known in sociolinguistic literature as langue de bois (wooden language)19. This language was characterized by rigid syntax, ideological clichés, and a detachment from observable reality. It functioned less to communicate empirical information than to assert the omnipresence of state power and enforce a singular, state-approved version of reality. In contemporary Russia, the manipulation of legal terminology serves as a powerful mechanism for suppressing civil society. The 2012 "Law on Foreign Agents" deliberately weaponized a highly stigmatized term to redefine public order and national security21. By requiring non-governmental organizations (NGOs) and media outlets receiving international funding to label themselves as "foreign agents," the state capitalized on the term's Soviet-era connotations of espionage, treason, and subversion21. This legislation effectively redefined domestic civic engagement, human rights advocacy, and independent journalism as hostile foreign influence21. The success of this semantic weaponization has led to its exportation across other authoritarian and illiberal governments. Regimes in Georgia, Kyrgyzstan, and Hungary have attempted to implement similar "foreign agent" or "transparency of foreign influence" laws21. Proponents of these laws frequently claim they are merely equivalents to the United States' Foreign Agents Registration Act (FARA)23. However, legal analysis reveals a critical semantic divergence: FARA targets entities acting under the direct control and direction of a foreign principal (a classic principal-agent relationship), whereas the Russian and Georgian models apply the stigmatizing label to independent organizations based solely on the origin of a fraction of their funding23. The strategic deployment of the term "foreign agent" creates a chilling effect, forcing self-censorship, draining resources through compliance, and isolating critics from public life without the immediate need for overt physical violence21. The European Court of Human Rights has repeatedly criticized such laws for creating an environment of forced self-stigmatization that bears the hallmarks of totalitarian control25.

Wartime Democracies: Propaganda and the Criminalization of Dissent

Democratic nations are not immune to the severe restriction of vocabulary, particularly during periods of existential threat. During World War I, the United States government undertook massive campaigns to control public discourse, redefine patriotism, and criminalize dissenting language under the banner of national security. President Woodrow Wilson established the Committee on Public Information (the Creel Committee), which executed a vast propaganda campaign to generate pro-war sentiment while managing a system of "voluntary" press censorship26. The government promoted a narrative of "100 percent Americanism," redefining civic loyalty as unquestioning support for the war effort and framing any opposition as dangerous extremism27. This environment led to the aggressive suppression of the German language; states prohibited teaching German, German-named institutions were rebranded, and citizens were subjected to vigilante violence for failing to exhibit sufficient linguistic loyalty27. Legally, this control was codified in the Espionage Act of 1917 and the Sedition Act of 191826. The Sedition Act explicitly criminalized the utterance, printing, writing, or publishing of any "disloyal, profane, scurrilous, or abusive language" about the U.S. government, the Constitution, or the military26. This legislation discarded the requirement to prove that speech caused direct, injurious consequences, allowing prosecutors to define virtually any anti-war sentiment as a threat to public order27. The semantic boundary of "treason" was thus expanded to encompass pacifism, labor activism, and political dissent, demonstrating how democratic states can rapidly deploy semantic manipulation to crush opposition26.

Corporate Euphemisms and the Reframing of Violence

In modern geopolitical conflicts and corporate public relations, institutions frequently rely on euphemisms to sanitize controversial actions, obscure moral implications, and manage public perception. Linguist Steven Pinker popularized the concept of the "euphemism treadmill," which describes the sociological process by which polite or clinical terms gradually accumulate the negative connotations of the concepts they describe, forcing society to continuously invent new euphemisms (a process of semantic drift)32. In corporate environments, terms regarding termination (e.g., "fired") drift into sanitized alternatives ("let go," "downsized," "right-sized"), reflecting an institutional desire to avoid linguistic friction35. However, in the political and military arena, euphemisms are utilized deliberately to construct a legal and moral shield around state violence37. During the "War on Terror," the United States government deployed the term "enhanced interrogation techniques" to describe practices that clearly met the international legal definitions of torture and inhuman treatment38. This deliberate lexical choice bypassed domestic and international prohibitions on torture by establishing a parallel, sanitized semantic category, effectively redefining extremism and state violence41. Similarly, the term "collateral damage" strips the emotional and moral weight from the death of non-combatants, reframing civilian casualties as a tragic but highly technical byproduct of legitimate military operations38. By speaking of conflicts "erupting" or "breaking out" like natural disasters, institutions utilize language that absolves human actors of responsibility39. Such language changes thought not by determining what people can physically imagine, but by directing their cognitive attention away from the visceral realities of violence toward abstract, bureaucratic justifications.

The Subversive Lexicon: Population Responses and Resistance

When institutional authorities restrict the public lexicon, populations do not become mute; rather, they adapt. Human agency asserts itself through the development of complex, coded linguistic strategies that allow individuals to communicate dissent, preserve identity, and critique power while evading detection.

Preference Falsification and the Public Transcript

Under strict censorship, individuals routinely engage in "preference falsification," a sociological concept articulated by Timur Kuran. Preference falsification occurs when people publicly express preferences or beliefs that contradict their private convictions, driven by the perceived threat of social or state retaliation43. When preference falsification becomes ubiquitous, it sustains social stability and conceals political fragility, creating "collective illusions" where the majority of a population complies with an ideology they privately despise because they incorrectly believe everyone else supports it43. The Czech dissident and playwright Václav Havel illustrated this dynamic in his seminal essay, The Power of the Powerless. Havel utilizes the metaphor of a greengrocer who places a sign reading "Workers of the world, unite\!" in his shop window47. The greengrocer does not display the slogan out of genuine ideological enthusiasm, but as a signal of compliance to avoid harassment by the authorities47. The slogan functions as a linguistic uniform, a ritualistic lie that contributes to the overall spiritual suffocation of the society47. However, Havel posits that if individuals cease to participate in these linguistic rituals—if they refuse to display the slogan and choose to "live in truth"—the entire facade of totalitarian power becomes vulnerable47.

Hidden Transcripts and Infrapolitics

The disparity between public compliance and private dissent is further explored by political scientist James C. Scott through the concept of the "hidden transcript"52. Scott divides discourse in highly unequal societies into two categories. The "public transcript" is the open, outward interaction between subordinates and those who dominate, typically characterized by deference, submission, and a misleading performance of acceptance52. Conversely, the "hidden transcript" is the discourse that occurs offstage, beyond the direct observation of power, encompassing jokes, gossip, songs, and coded conversations that critique and undermine authority52. Cultural resistance frequently relies on the hidden transcript to survive in plain sight. This resistance does not require principled, overt rebellion; it manifests as "infrapolitics"—the subtle, everyday acts of evasion and subversion55. For example, in ancient Indian texts, subversive philosophical ideas were sometimes smuggled into orthodox religious documents by attributing them to a "prior faction" or a theoretical opponent, allowing the author to present forbidden thoughts under the guise of refuting them57. Similarly, the lyrics of blind Blues singers in the American South operated as a hidden transcript, conveying deep social critique and survival strategies through aesthetic and musical codes illegible to the dominant racial hierarchy53.

Aesopian Language and Coded Digital References

In contexts of systemic state censorship, such as Tsarist and Soviet Russia, writers developed a highly sophisticated literary system known as "Aesopian language," thoroughly analyzed by the scholar Lev Loseff58. Aesopian language relies on allegory, irony, circumlocution, and obscure cultural references to transmit forbidden political messages to the reader while maintaining a superficial veneer of innocence to bypass the state censor58. This dynamic creates a triangular relationship between the author, the reader, and the censor, where the reader must actively decode the text to understand the political subtext58. A contemporary analogue to Aesopian language can be found in Chinese online communities, where internet users constantly develop slang, memes, and intentional misspellings to bypass state firewalls. Authors utilize the ancient zhiguai (ghost story) genre to construct allegorical political commentary. By veiling their critiques of modern governance in bizarre, supernatural narratives, writers create a dual-layered text that evades algorithmic detection while delivering the hidden transcript to an initiated audience62. When automated systems detect specific keywords, populations respond with homophones, pinyin acronyms, and image-based memes, turning linguistic evasion into a cultural art form.

Algorithmic Censorship and the Digital Panopticon

In the contemporary era, the primary arbiters of public discourse are no longer solely state censors reading physical manuscripts, but corporate social media platforms relying on automated, algorithmic content moderation. To avoid demonetization, shadowbanning, or account deletion, digital populations have developed rapidly evolving lexicons designed specifically to defeat machine learning models.

Dynamic Keyword Filtering and OCR Censorship

Authoritarian states possess the capability to mandate algorithmic censorship across domestic technology platforms. Research by Citizen Lab on Chinese platforms like WeChat reveals the immense scale of this semantic control. Censors utilize dynamic, constantly updated keyword lists to detect and block politically sensitive content in real-time63. As users attempt to evade text-based filters by embedding forbidden terminology into images or memes, authorities deploy Optical Character Recognition (OCR) technology to scan, decode, and censor text within images63. This creates an arms race between automated detection and human creativity, where citizens must continuously distort language—using rotated text, obscure visual metaphors, and complex intentional misspellings—to communicate about redefined concepts like public order and misinformation.

The Mechanics of Algospeak

Even in democratic nations, corporate platforms enforce strict vocabulary guidelines, giving rise to "algospeak." Algospeak represents a direct, evolutionary response to algorithmic content moderation, functioning as a high-speed game of linguistic Whac-A-Mole32. When algorithms are programmed to detect and suppress specific keywords related to sexuality, violence, mental health, or controversial politics, users invent morphological, phonetic, and semantic workarounds32. Prominent examples, particularly popularized on platforms like TikTok, include the use of "unalive" or "sewer slide" instead of "suicide," "seggs" for "sex," and "le$bean" for "lesbian"32. Emojis are frequently weaponized for acoustic or visual substitution, such as the corn emoji for "porn," the eggplant emoji for genitalia, or the ninja emoji to replace racial slurs32. Unlike earlier iterations of internet evasion, such as "leetspeak" (substituting numbers for letters, like 5U1C1D3), modern algospeak operates against sophisticated, AI-driven natural language processing systems, requiring constant, creative, and culturally embedded mutation32.

Majority Understandable Modulation (MUM)

The necessity of utilizing algospeak creates a complex communicative trade-off for digital citizens. Researchers have formalized this dynamic through the concept of Majority Understandable Modulation (MUM)64. When users modulate their language to evade algorithmic detection, they simultaneously increase the cognitive load required for their human audience to decipher the message. The MUM point is defined as the exact threshold of linguistic alteration at which the text successfully evades the algorithmic detector, but beyond which it begins to lose comprehension for the majority of the intended human audience64. If the modulation exceeds the MUM point, the language ceases to be a tool of public communication and devolves into entirely closed, coded language64. Furthermore, the pervasive use of algospeak highlights the failure of non-contextual moderation systems, which frequently penalize marginalized communities66. For instance, LGBTQ+ creators utilizing platforms for community support are forced to censor their own identities to avoid being incorrectly categorized as violating safety guidelines66. This algorithmic erasure echoes historical forms of censorship, where the vocabulary necessary to articulate a specific social reality is systematically suppressed by the governing architecture.

Generative AI: The New Frontier of Vocabulary Control

The rapid proliferation of Large Language Models (LLMs) represents the most sophisticated evolution of automated language control. Because generative AI systems serve as intermediaries between users and vast repositories of human knowledge, the mechanisms governing their outputs effectively dictate the boundaries of accessible digital reality68. Generative AI can create a new form of vocabulary control in which models systematically avoid, reframe, or substitute politically sensitive terminology, shifting from reactive filtering to proactive narrative generation.

State-Mandated Alignment: China's Interim Measures

Governments have quickly recognized the ideological power of generative AI and sought to impose strict regulatory frameworks to control its vocabulary. In August 2023, the Cyberspace Administration of China (CAC) enacted the Interim Measures for the Management of Generative Artificial Intelligence Services, marking the world's first comprehensive, legally binding regulation specifically targeting public-facing generative AI70. Unlike Western frameworks that predominantly focus on data privacy, copyright, or high-risk systemic failures, the Chinese regulatory regime is profoundly ideological74. The Interim Measures mandate that all generative AI outputs must "uphold socialist core values"70. Models are strictly prohibited from generating content that incites subversion of state power, endangers national security, or damages the national image—broad categories historically utilized to suppress political dissent70. Providers of AI services must undergo mandatory security assessments and algorithm filings before launch, ensuring that the model's vocabulary and conceptual mapping align flawlessly with state orthodoxy72. If a model generates prohibited content, the provider must promptly filter the output, update the training data, and report the incident to authorities, effectively integrating the AI into the state's apparatus of surveillance and censorship70.

Soft versus Hard AI Censorship

Outside of strictly regulated authoritarian environments, commercial LLM providers worldwide implement extensive safety fine-tuning (e.g., Reinforcement Learning from Human Feedback, or RLHF, and Direct Preference Optimization, DPO) to prevent the generation of toxic, illegal, or biased content69. However, these safety mechanisms frequently result in the systematic censorship of legitimate political discourse. Researchers evaluating AI behavior classify LLM censorship into two distinct manifestations:

1. Hard Censorship: The model explicitly refuses to answer a prompt, often generating a canned response (e.g., "As an AI language model, I cannot provide..."), an error message, or a completely off-topic deflection68.

2. Soft Censorship: The model provides a compliant response but selectively omits, downplays, or slants key factual elements regarding sensitive topics, thereby reframing the reality of the situation without explicitly refusing the prompt68.

Evaluations of leading LLMs utilizing datasets like the Do-Not-Answer benchmark reveal that models frequently exhibit ideological biases tied to their corporate or geographic origins68. For example, models may engage in extreme soft censorship regarding domestic political figures or historical events (such as the Tiananmen Square protests or the treatment of Uyghurs in Xinjiang by Chinese-developed models), selectively suppressing facts while amplifying state-aligned language69. The application of safety guardrails often leads to "over-refusal," where benign queries are rejected merely because they contain terms adjacent to forbidden topics81.

Technical Mechanisms of Suppression and Transparency

The technical implementation of AI censorship operates at both the model representation level and the inference (generation) level, essentially encoding forbidden vocabulary directly into the mathematics of the neural network. To counter this, researchers have developed transparency mechanisms to reveal and manipulate when an AI's vocabulary is constrained. At the final stage of text generation, LLMs select the next word based on a probability distribution over their entire vocabulary (the logits). AI platforms frequently employ "logit processors" to mathematically suppress specific tokens from ever being generated82. During the decoding process, a logit processor can intercept the probability scores and apply a negative infinity mask to forbidden words, ensuring the model is physically incapable of outputting them, regardless of the prompt's context82. While this is an effective tool for preventing the output of slurs, it constitutes a brute-force approach to vocabulary control that operates entirely outside the model's semantic understanding84. A more profound transparency and control mechanism is found in "representation engineering" (RepE)76. Instead of focusing on individual neurons or output tokens, RepE identifies how high-level concepts (such as honesty, harmfulness, or refusal) are represented globally across the network's internal activations76. By analyzing contrasting inputs (e.g., a benign prompt versus a politically sensitive prompt), researchers can locate a specific "refusal-compliance vector"76. Once this directional vector is identified, developers can utilize "activation steering" at inference time to manipulate the model's behavior76. By artificially adding or subtracting this vector from the model's internal activations, researchers can force an otherwise compliant model to strictly censor its output, or conversely, bypass the model's safety training to reveal the political truths it was trained to suppress69. This mechanism demonstrates that censorship in modern AI is not merely a list of banned words, but a geometric constraint within the model's latent space, and proves that models often possess factual knowledge they are explicitly programmed to hide69.

Conclusion: Preserving Linguistic Diversity and Intellectual Freedom

The history of language as a tool of censorship demonstrates a fundamental political reality: those who control the vocabulary control the parameters of acceptable thought. Whether through the brutal linguistic engineering of Maoist China and the Third Reich, the bureaucratic euphemisms of modern democracies, or the algorithmic filtering of social media and generative AI, the underlying mechanism remains consistent. Restricting the lexicon limits the ability of a population to articulate dissent, perceive nuance, and challenge institutional power. Language does not completely determine thought, but semantic manipulation actively shapes cognitive focus, moral judgment, and political action. However, the resilience of human communication is equally apparent. The perpetual emergence of hidden transcripts, Aesopian literature, intentional misspellings, and algospeak proves that populations will continually invent new linguistic pathways to circumvent top-down semantic constraints. Language is fundamentally dynamic, inherently resisting permanent capture by rigid ideological frameworks. As the dissemination of information becomes increasingly mediated by artificial intelligence and algorithmic platforms, the preservation of linguistic diversity and intellectual freedom faces unprecedented systemic challenges. Generative models possess the capacity to enforce soft censorship at a global scale, subtly reshaping historical and political narratives without the transparency of traditional, state-sponsored censorship mechanisms. To mitigate this threat while recognizing the legitimate necessity of content moderation (such as preventing the dissemination of illegal material or genuine incitement to violence), the following principles must be established:

1. Algorithmic Transparency and Auditing: The specific mechanisms used to align and censor AI models—including training data exclusions, activation steering directions, and logit penalties—must be subject to independent audit. Users must be explicitly informed when a model's outputs are constrained by state regulation, corporate policy, or safety fine-tuning.

2. Geographic and Ideological Plurality: Reliance on a monopolistic ecosystem of LLMs centralizes semantic control. Promoting open-weight models from diverse cultural and geographic backgrounds is essential to prevent a singular, globally enforced standard of soft censorship.

3. Contextual Moderation over Lexical Bans: Both social media algorithms and AI systems must evolve beyond simple keyword suppression. True safety requires context-aware moderation that protects marginalized communities from harm without penalizing the vocabulary they use to describe their own experiences or stifling academic and political discourse.

Protecting the diversity of the lexicon is not merely a matter of academic or cultural preservation; it is the fundamental prerequisite for sustaining the cognitive liberty required for a free and democratic society.

Works cited

1. Linguistic Engineering: Language and Politics in Mao's China (review), https://muse.jhu.edu/article/184403/summary

2. inguistic \- ScholarSpace, https://scholarspace.manoa.hawaii.edu/bitstreams/b37cf237-75d2-426b-abad-a527d98c222d/download

3. PIERRE BOURDIEU AND THE PRACTICES OF LANGUAGE, https://www.annualreviews.org/content/journals/10.1146/annurev.anthro.33.070203.143907

4. 10 Fields, Trajectories, and Symbolic Power: Studying Practices of, https://academic.oup.com/book/46568/chapter/408132457

5. Introduction \- Language as Symbolic Power, https://www.cambridge.org/core/books/language-as-symbolic-power/introduction/EADD5DAC6206719402B67568A245495F

6. LTI – Lingua Tertii Imperii \- Wikipedia, https://en.wikipedia.org/wiki/LTI\_%E2%80%93\_Lingua\_Tertii\_Imperii

7. WHAT THE LANGUAGE OF THE THIRD REICH \- Re-UNIR, https://reunir.unir.net/bitstreams/ec80b798-922f-4177-ae4b-f4f42a1876da/download

8. Recalling Victor Klemperer (1881-1960) and his 'Language of the, https://www.peoplesworld.org/article/recalling-victor-klemperer-1881-1960-and-his-language-of-the-third-reich/

9. The Language of the Third Reich | Jörg Riecke \- Inference Review, https://inference-review.com/article/the-language-of-the-third-reich

10. Victor Klemperer's Language-Critical Reflections on Anti-Semitic, https://resolve.cambridge.org/core/services/aop-cambridge-core/content/view/526B3086A155008715534FD91AA2E89D/9781782044642c6\_p89-104\_CBO.pdf/stigma\_and\_performance\_victor\_klemperers\_languagecritical\_reflections\_on\_antisemitic\_hate\_speech.pdf

11. Victor Klemperer and the Decay of Political Language, https://www.the-american-interest.com/2020/01/12/victor-klemperer-and-the-decay-of-political-language/

12. LABELS AND THEIR CONSEQUENCES IN MAO'S CHINA (1949, https://www.openstarts.units.it/bitstreams/ce55d92e-4654-44e1-9b6d-a9c9e0820a4e/download

13. Linguistic Engineering : Language and Politics in Mao's China, https://onesearch.library.northeastern.edu/discovery/fulldisplay/alma9951756964301401/01NEU\_INST:NUL

14. Cultural Revolution, Effect on Language \- Brill \- Reference Works, https://referenceworks.brill.com/display/entries/ECLO/COM-00000113.xml

15. Common Speech: How China Invented Their National Language, https://sjquillen.medium.com/common-speech-how-china-invented-their-national-language-f5b9b6b1910a

16. The People's Language (Chapter 4\) \- Dialect and Nationalism in, https://www.cambridge.org/core/books/dialect-and-nationalism-in-china-18601960/peoples-language/498077B243D4D586199CC206F9B592B9

17. Political Ideology and Language Policy in North Korea \- ResearchGate, https://www.researchgate.net/publication/290511966\_Political\_Ideology\_and\_Language\_Policy\_in\_North\_Korea

18. State Ideology And Language Policy In North Korea \- ScholarSpace, https://scholarspace.manoa.hawaii.edu/items/a5717b41-5b13-484e-bef3-9e8ec0578ad6

19. June 2021 \- The Brazen Head, https://brazen-head.org/2021/06/

20. Histoire de la communication mondiale | PDF \- Scribd, https://fr.scribd.com/document/785217661/Sciences-Humaines-Et-Sociales-Armand-Mattelart-La-Communication-monde-Histoire-Des-Idees-Et-Des-Strategies-La-Decouverte-Poche-1994

21. Foreign Agent Laws in the Authoritarian Playbook, https://www.hrw.org/news/2024/09/19/foreign-agent-laws-authoritarian-playbook

22. Full article: International actors and democracy protection, https://www.tandfonline.com/doi/full/10.1080/13510347.2025.2461463

23. Georgia's new law on “transparency of foreign influence” and its, https://fpc.org.uk/georgias-new-law-on-transparency-of-foreign-influence-and-its-incompatibility-with-international-human-rights-standards/

24. How Georgia's foreign agents law differs from FARA \- JAMnews, https://jam-news.net/how-us-fara-differs-from-georgias-foreign-agents-law-lawyer-explains/

25. Georgia's Foreign Agent Law 2.0 \- Verfassungsblog, https://verfassungsblog.de/georgias-foreign-agent-law-2-0-ecthr/

26. Mary Dudziak on Civil Liberties During WWI and Beyond, https://www.carnegiecouncil.org/media/series/100/mary-dudziak-on-civil-liberties-during-wwi-and-beyond

27. Americans Toss Lady Liberty Overboard During Crises | Cato Institute, https://www.cato.org/commentary/americans-toss-lady-liberty-overboard-during-crises

28. U.S. Curtails Civil Liberties During World War I | History \- EBSCO, https://www.ebsco.com/research-starters/history/us-curtails-civil-liberties-during-world-war-i/

29. Propaganda and Civil Liberties During World War I | History \- EBSCO, https://www.ebsco.com/research-starters/history/propaganda-and-civil-liberties-during-world-war-i

30. Stopping the Hun: Anti-German sentiment in World War I, https://www.researchgate.net/publication/305766906\_Stopping\_the\_Hun\_Anti-German\_sentiment\_in\_World\_War\_I

31. APUSH Chapter 31, War to End War \- American History Central, https://www.americanhistorycentral.com/entries/progressive-era-woodrow-wilson-and-world-war-i/

32. Algospeak – Complete Book Summary & All Key Ideas \- HowToes, https://howtoes.blog/2025/07/28/algospeak-complete-book-summary-all-key-ideas/

33. 39: Is This a Reference? (with Sylvia Sierra) \- Because Language, https://becauselanguage.com/39-is-this-a-reference/

34. What's the name for this social process? Example: “homeless, https://www.reddit.com/r/linguistics/comments/142vkx5/whats\_the\_name\_for\_this\_social\_process\_example/

35. Anna Feldman's research works | Montclair State University, https://www.researchgate.net/scientific-contributions/Anna-Feldman-69705590

36. Why is it accepted to call someone a " lunatic " and its not ... \- Reddit, https://www.reddit.com/r/stupidquestions/comments/1h4qbnu/why\_is\_it\_accepted\_to\_call\_someone\_a\_lunatic\_and/

37. How Euphemisms Are Used in the Political Arena \- ResearchGate, https://www.researchgate.net/publication/265098807\_Truths\_and\_Euphemisms\_How\_Euphemisms\_Are\_Used\_in\_the\_Political\_Arena

38. EUPHEMISMS IN THE ENGLISH MASS MEDIA \- inLibrary, https://inlibrary.uz/index.php/ejar/article/download/139573/140433/202801

39. The Wars of the Worlds and the Politics of Destruction \- Medium, https://medium.com/@krigerbruce/the-wars-of-the-worlds-and-the-politics-of-destruction-0d8d94509768

40. Necropolitical Law (Chapter One) \- Discounting Life, https://www.cambridge.org/core/books/discounting-life/necropolitical-law/C6326519FCA7EAA04DF05F4B3FE7BB1B

41. Doc. 11302 \- Report \- Working document, https://pace.coe.int/files/11555/html

42. The Gospel of Guantánamo | Harvard Divinity Bulletin, https://bulletin.hds.harvard.edu/the-gospel-of-guantanamo/

43. Preference falsification \- Simple English Wikipedia, the free, https://simple.wikipedia.org/wiki/Preference\_falsification

44. Preference Falsification: Making Sense of Public Opinion Surveys in, https://www.tandfonline.com/doi/full/10.1080/03068374.2025.2519632

45. Preference Falsification And Cascade \- OpEd \- Eurasia Review, https://www.eurasiareview.com/19122024-preference-falsification-and-cascade-oped/

46. Full article: The autocratic bias: self-censorship of regime support, https://www.tandfonline.com/doi/full/10.1080/13510347.2021.1981867

47. The Power of the Powerless (essay) | CNCR \- Nonviolent-conflict.org, https://www.nonviolent-conflict.org/resource/the-power-of-the-powerless/

48. Ideology and the Corruption of Language \- Public Discourse, https://www.thepublicdiscourse.com/2017/03/18571/

49. “The Power of the Powerless” by Vaclav Havel \- Medium, https://bruces.medium.com/the-power-of-the-powerless-by-vaclav-havel-84b2b8d3a84a

50. Remembering Vaclav Havel | The View East \- WordPress.com, https://thevieweast.wordpress.com/2011/12/23/remembering-vaclav-havel/

51. "The Power of the Powerless" \- Vaclav Havel, https://hac.bard.edu/amor-mundi/the-power-of-the-powerless-vaclav-havel-2011-12-23

52. Hidden Transcripts? The Supposedly Self-Censoring Paul and, https://www.cambridge.org/core/journals/new-testament-studies/article/hidden-transcripts-the-supposedly-selfcensoring-paul-and-rome-as-surveillance-state-in-modern-pauline-scholarship/708DB6C63356DAAAFEE49C80274410A1

53. Hidden Rhythms: Cultural Resistance in a New Generation of Artists, https://www.ourppls.com/post/hidden-rhythms-cultural-resistance-in-a-new-generation-of-artists

54. Subaltern Readings of Blues Music and Lyrics \- Journals@KU, https://journals.ku.edu/africanaannual/article/download/21870/21684/89630

55. Full article: Everyday forms of resistance in theory and practice, https://www.tandfonline.com/doi/full/10.1080/03066150.2026.2630677

56. infrapolitics | Postsocialism, https://postsocialism.org/tag/infrapolitics/

57. Prelude to Censorship: The Toleration of Blasphemy in Ancient India, https://divinity.uchicago.edu/sightings/articles/prelude-censorship-toleration-blasphemy-ancient-india

58. (Loseff Lev) On The Beneficence of Censorship. Aes PDF \- Scribd, https://de.scribd.com/document/454179154/Loseff-Lev-On-The-Beneficence-Of-Censorship-Aes-z-lib-org-pdf

59. On the Beneficence of Censorship: Aesopian Language in Modern, [https://www.cambridge.org/core/journals/slavic-review/article/on-the-beneficence-of-censorship-aesopian-language-in-modern-russian-literature-by-lev-loseff-translated-by-jane-bobko-arbeiten-und-texte-zur-slavistik-vol-31-edited-by-wolfgang-kasack-munich-verlag-otto-sanger-1984-xii-274-pp-tables-280-s-paper/08ABB3C2FBB51CA0924B8D5C1AE1B9C3](https://www.cambridge.org/core/journals/slavic-review/article/on-the-beneficence-of-censorship-aesopian-language-in-modern-russian-literature-by-lev-loseff-translated-by-jane-bobko-arbeiten-und-texte-zur-slavistik-vol-31-edited-by-wolfgang-kasack-munich-verlag-otto-sanger-1984-xii-274-pp-tables-280-s-paper/08ABB3C2FBB51CA0924B8D5C1AE1B9C3)

60. Reviews 399 \- Cambridge University Press & Assessment, [https://www.cambridge.org/core/services/aop-cambridge-core/content/view/08ABB3C2FBB51CA0924B8D5C1AE1B9C3/S0037677900106552a.pdf/on-the-beneficence-of-censorship-aesopian-language-in-modern-russian-literature-by-lev-loseff-translated-by-jane-bobko-arbeiten-und-texte-zur-slavistik-vol-31-edited-by-wolfgang-kasack-munich-verlag-o.pdf](https://www.cambridge.org/core/services/aop-cambridge-core/content/view/08ABB3C2FBB51CA0924B8D5C1AE1B9C3/S0037677900106552a.pdf/on-the-beneficence-of-censorship-aesopian-language-in-modern-russian-literature-by-lev-loseff-translated-by-jane-bobko-arbeiten-und-texte-zur-slavistik-vol-31-edited-by-wolfgang-kasack-munich-verlag-o.pdf)

61. On the Beneficence of Censorship: Aesopian Language in Modern, https://books.google.com/books/about/On\_the\_Beneficence\_of\_Censorship.html?id=E-J9AAAAIAAJ

62. Contemporary Chinese Online Allegorical Ghost Stories as Political, https://www.metacriticjournal.com/article/117/new-wine-in-old-bottles-contemporary-chinese-online-allegorical-ghost-stories-as-political-commentary

63. Q\&A: Citizen Lab documents Chinese censorship of coronavirus, https://cpj.org/2020/03/citizen-lab-chinese-censorship-coronavirus/

64. Algospeak, Hiding in the Open: The Trade-off Between Legible, https://arxiv.org/pdf/2605.06619

65. The “algospeak” dialect \- Cory Doctorow \- Medium, https://doctorow.medium.com/the-algospeak-dialect-74961b4803b7

66. Using Algospeak to Contest and Evade Algorithmic Content, https://www.researchgate.net/publication/374413909\_You\_Can\_Not\_Say\_What\_You\_Want\_Using\_Algospeak\_to\_Contest\_and\_Evade\_Algorithmic\_Content\_Moderation\_on\_TikTok

67. Using Algospeak to Contest and Evade Algorithmic Content, https://par.nsf.gov/biblio/10480449-you-can-say-what-you-want-using-algospeak-contest-evade-algorithmic-content-moderation-tiktok

68. arXiv:2504.03803v1 \[cs.CL\] 4 Apr 2025, https://arxiv.org/pdf/2504.03803

69. Fine-grained Control over LLM Refusal Behaviour for Sensitive Topics, https://arxiv.org/pdf/2512.16602

70. China's 2023 Generative AI Regulations | PDF | Artificial Intelligence, https://www.scribd.com/document/877586334/AI-Watch-Global-Regulatory-Tracker-China-White-Case-LLP

71. Navigating China's regulatory approach to generative artificial, https://www.cambridge.org/core/services/aop-cambridge-core/content/view/969B2055997BF42DE693B7A1A1B4E8BA/S3033373324000048a.pdf/navigating-chinas-regulatory-approach-to-generative-artificial-intelligence-and-large-language-models.pdf

72. Emergency Response Measures for Catastrophic AI Risk \- arXiv, https://arxiv.org/html/2511.05526v1

73. China AI Regulation \- Deep Lex, https://www.deep-lex.com/ai-regulation-tracker/china

74. China AI Regulations 2026: Rules Companies Must Follow, https://www.pertamapartners.com/insights/china-ai-regulations

75. Making Sense of China's AI Regulations \- Holistic AI, https://www.holisticai.com/blog/china-ai-regulation

76. Steering the CensorShip: Uncovering Representation Vectors for, https://arxiv.org/html/2504.17130v3

77. Steering the CensorShip: Uncovering Representation Vectors for, https://arxiv.org/html/2504.17130v1

78. An Empirical Study of Moderation and Censorship Practices \- arXiv, https://arxiv.org/html/2504.03803v1

79. arXiv:2308.13387v2 \[cs.CL\] 4 Sep 2023, https://arxiv.org/pdf/2308.13387

80. Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation, https://arxiv.org/html/2603.05494v2

81. CSSBench: Evaluating the Safety of Lightweight LLMs ... \- arXiv, https://arxiv.org/pdf/2601.00588

82. Utilities for Generation \- Hugging Face, https://huggingface.co/docs/transformers/internal/generation\_utils

83. Large Language Models \- Ludwig, https://ludwig.ai/latest/configuration/large\_language\_model/

84. A Versatile Plug-in for Handling Negative Constraints in Decoding, https://arxiv.org/pdf/2605.10065

85. Uncovering Control-Plane Vulnerabilities in LLMs with Structured, https://arxiv.org/pdf/2503.24191

86. Representation Engineering for Large-Language Models \- arXiv, https://arxiv.org/html/2502.17601v1