The Semantic Weaponization

The Semantic Record series (SR-001 through SR-006) documents three independent mechanisms by which the shared, reliable meaning of a word is corrupted. Each attacks a different component of the same target: the linguistic infrastructure on which cognition, regulation, and democratic deliberation depend.

The first mechanism is the Euphemism Treadmill, documented in SR-001. It attacks the signifier — the word itself. A term that functions as an embedded instruction — triggering regulatory response, clinical action, or public alarm — is replaced by a substitute term that disables the instruction while preserving the appearance of language functioning normally. “Addiction” becomes “pseudo-addiction.” “Surveillance” becomes “personalized experience.” “Censorship” becomes “content moderation.” In each case, the phenomenon remains identical. Only the word changes. And with the word, the response it would have triggered is silently removed. SR-001 names this condition Semantic Inversion: “the deliberate replacement of a term that signals harm with a term that signals benefit (or signals nothing), where the new term is operationally designed to prevent the regulatory, clinical, or behavioural response that the original term would have triggered.”

The second mechanism is Tripwire Relocation, documented in SR-002. It attacks the definition — what the word officially means. This is more sophisticated than the treadmill: the word stays the same, but its operational boundary is moved. The NSA’s redefinition of “collection” is the paradigmatic case. Under the Foreign Intelligence Surveillance Act of 1978, acquisition constituted collection, which triggered Fourth Amendment protections. After 2001, the intelligence community adopted a new definition: “Data is not ‘collected’ when it is ingested by a computer system and stored in a database. Data is ‘collected’ only when a human analyst queries the database and views the result.” The statute remained intact. The word remained the same. The definition was moved to a point downstream of the behavior it was designed to catch — allowing the bulk surveillance of every American’s communications metadata while officials testified truthfully, under their own definitional framework, that no “collection” was occurring. SR-002 names this condition The Definitional Bypass.

The third mechanism is Gravity Dilution, documented in SR-003. It attacks the scope — what the word applies to. A high-stakes term is progressively expanded to cover lower-stakes phenomena. The prescriptive force of the term is preserved (initially), but its range of application expands until the original targets are lost in the noise. “Domestic terrorist” expands from Oklahoma City bombers to environmental protesters to pipeline activists to journalists covering protests. “Trauma” expands from events outside the range of usual human experience (DSM-III, 1980) to workplace conflict and microaggressions. The term’s gravity is diluted bidirectionally: it is over-deployed against the expanded targets while being functionally weakened for the original targets. SR-003 names this condition Concept Creep.

The three mechanisms attack the same target — the shared, reliable meaning of a word — through different vectors. Together, they can render any regulatory, clinical, or civic term structurally unreliable. A term can be replaced (treadmill), redefined (relocation), or diluted (gravity) — or, as the Purdue case demonstrates, all three in sequence on a single product. The unified framework is CV-020’s first original contribution: the SR series documented each mechanism individually. This paper shows they constitute a coordinated system.

The Purdue Pharma Semantic Record (SR-004) documents all three mechanisms operating in sequence on a single product. The case is the proof of convergence — the forensic demonstration that semantic capture is not a metaphor for institutional failure but the mechanism of it.

1989 — Euphemism Treadmill. J. David Haddox published a single paragraph in the journal Pain coining the term “pseudo-addiction.” The letter contained no controlled data, no prospective study, no clinical trial. It was not a peer-reviewed research article. It was a letter — and it introduced a term that would be cited in medical textbooks, embedded in continuing medical education programmes, deployed in sales training materials, and used to convince physicians that patients displaying drug-seeking behavior were not addicted but undertreated. By 1996, “pseudo-addiction” had achieved the status of established clinical knowledge. The euphemism treadmill had replaced “addiction” with a term that prescribed the opposite clinical response: not withdrawal of the drug, but escalation.

1996 — Tripwire Relocation. OxyContin’s FDA label was expanded from “Schedule II narcotic with significant abuse potential” to include “appropriate for moderate to severe pain, including chronic non-cancer pain” — at Purdue’s request, supported by Purdue-funded data. The word “appropriate” had not changed. The boundary of what it covered had been moved — from acute post-surgical pain to the vast, chronic, and diagnostically ambiguous category of non-cancer pain. The definitional tripwire had been relocated to a point downstream of the population the drug would now reach.

2001–2010 — Gravity Dilution. The category “pain patient” was progressively expanded to include populations whose risk profiles were internally documented as high. The term’s prescriptive force — “pain patients deserve adequate treatment” — was preserved. Its scope was expanded until it covered populations for whom opioid prescribing carried documented risks that Purdue’s own scientists acknowledged.

Purdue’s own scientists were aware that the distinction between addiction and “pseudo-addiction” was clinically unverifiable in practice. The two conditions presented identically: drug-seeking behavior, dose escalation requests, preoccupation with obtaining the medication. The recommended differential diagnosis — “escalate the opioid and observe whether behavior normalizes” — was operationally indistinguishable from fueling addiction. The euphemism did not merely obscure the diagnosis. It prescribed the action that worsened the condition.

What the Purdue case proves: semantic capture is the mechanism of institutional failure, not a description of it. Five hundred thousand opioid deaths trace to a cascade that began with the corruption of three words — “addiction,” “appropriate,” “patient” — through three sequential mechanisms. The regulatory system remained intact throughout. The words it relied on had been emptied of their meaning.

The Counter-Semantic Standard (SR-006) establishes the methodology for distinguishing semantic capture from natural linguistic evolution. Not all meaning change is capture. Languages evolve. The word “computer” once meant a person who computes; now it means a machine. That shift was driven by technological change, not institutional interest. The question is not whether a word’s meaning has changed, but who benefited from the change and how the change was initiated.

SR-006 establishes three forensic criteria:

Criterion 1: Beneficiary Initiation. “Was the definitional change initiated, funded, or drafted by the entity that benefits from the change?” Purdue Pharma coined “pseudo-addiction” and funded its dissemination. The NSA drafted the redefinition of “collection” that exempted its surveillance programmes. The tobacco industry designed “light” cigarette categories that its own research showed did not reduce harm. In each case, the beneficiary of the meaning change was the author of the meaning change.

Criterion 2: Downstream Displacement. “Did the definitional change move a regulatory, clinical, or behavioural tripwire to a point downstream of the behaviour the tripwire was designed to catch?” The “collection” redefinition moved the Fourth Amendment trigger downstream of data ingestion. The “natural” non-definition left the trigger undefined entirely. OxyContin’s label expansion moved the prescribing boundary downstream of populations with documented risk profiles.

Criterion 3: Descriptive Displacement. “Does the pre-change term still accurately describe the phenomenon the post-change term was substituted for?” “Addiction” still accurately described the condition “pseudo-addiction” was invented to rename. “Surveillance” still accurately describes what “personalized experience” was deployed to obscure. “Censorship” still accurately describes the editorial function that “content moderation” was selected to euphemize.

The three criteria together establish a forensic methodology that is falsifiable, applicable across domains, and resistant to the most common objection to semantic capture analysis: that language simply evolves and any attempt to identify “corrupted” meanings is itself an exercise in authority. The SR-006 framework does not claim authority over meaning. It asks whether the entity that benefits from a meaning change is the entity that initiated the change. The question is evidentiary, not ideological.

The companion framework is SA-003’s Authority Laundering analysis, which documents how institutional interests are transmitted through the currency of ultimate authority — divine will, scientific consensus, democratic mandate — and emerge as legitimized power that cannot be challenged without appearing to challenge the authority itself. Authority laundering corrupts the source of a claim’s legitimacy. Semantic capture corrupts the words in which the claim is made. Same mechanism, different substrate. Together, the institution borrows unchallengeable authority and deploys it through captured vocabulary.

Semantic capture without distribution is a document in a filing cabinet. The Influence Architecture series (IA-001 through IA-006) documents the three mechanisms by which captured language achieves population-scale adoption.

Affective Engineering (IA-001) is the first distribution mechanism. The emotion is the distribution mechanism; the content is the payload. Facebook’s internal design assigned the “angry” reaction emoji a weight five times higher than “like” in its engagement scoring algorithm — a design decision by an engineering team that understood outrage as a feature, not a bug, of the distribution architecture. The amygdala processes threat-related stimuli approximately 100 milliseconds faster than the prefrontal cortex can evaluate them. Content designed to trigger emotional response — outrage, fear, moral disgust — achieves distribution before analytical processing can evaluate whether a term means what the audience assumes. A corrupted term embedded in affectively engineered content bypasses the cognitive defenses that would detect the corruption under conditions of reflective evaluation.

Consensus Engineering (IA-002) is the second distribution mechanism. The Oxford Computational Propaganda Project documented organized social media manipulation in 81 countries by 2020, up from 28 in 2017. Meta removed more than 200 coordinated inauthentic behavior operations across 70-plus countries. The Internet Research Agency deployed hundreds of operatives managing thousands of fictitious American personas at a budget of approximately $1.25 million per month. The mechanism exploits the social proof heuristic: humans calibrate their beliefs partly by reference to what others appear to believe. Synthetic consensus — the manufactured appearance of distributed organic agreement — provides the social proof that captured vocabulary is universally accepted. As IA-002 states: “Semantic capture provides the vocabulary; consensus engineering provides the social proof that the vocabulary is universally accepted.”

Source Laundering (IA-003) is the third distribution mechanism. The structural impossibility, for the end-user, of tracing information content to its actual origin, funder, or strategic intent. The tobacco industry’s Tobacco Industry Research Committee (TIRC) is the paradigmatic case: industry-funded research laundered through academic institutions, published in peer-reviewed journals, cited by subsequent researchers who had no knowledge of the funding source. The Purdue KOL (key opinion leader) network, documented in OA-002, operated identically: Purdue-funded physicians published research, delivered CME presentations, and authored textbook chapters that embedded “pseudo-addiction” as clinical science without disclosing the relationship. IA-003 identifies the paradox: “The more effectively a population is trained to evaluate sources, the more valuable laundering becomes — laundered sources survive evaluation.”

Captured language that is distributed but not monetized is advocacy. The Narrative Market series (NM-001 through NM-005) documents the three mechanisms by which corrupted meaning is converted into financial and political power — and why this conversion makes semantic capture structurally stable.

The Platform Authority Premium (NM-001) documents the financial value embedded in an individual’s platform authority that allows public statements to function as capital instruments. Jensen Huang’s AI projections moved Nvidia from $300 billion to approximately $3.5 trillion. Elon Musk’s single-word tweet “Gamestonk!!” moved GameStop’s stock price more than 100% in a single day. His Dogecoin endorsements moved the token from $0.05 to $0.73. NM-001’s load-bearing observation: “The insider trading framework protects markets against people who know things. It has no answer for people who are things.” When an individual’s words are the market event, the distinction between information and influence collapses — and the language in which markets operate becomes directly monetizable.

The Position-Before-Signal (NM-004) documents narrative arbitrage: establishing financial positions before generating the narrative signal that moves prices. Chamath Palihapitiya’s SPAC empire extracted approximately $750 million from six transactions while retail investors absorbed the losses — Virgin Galactic from $46 to $5, Clover Health from $28 to $4. The SEC’s December 2022 enforcement against eight social media influencers documented approximately $100 million in illicit profits from coordinated pump-and-dump schemes. The arbitrage operates not between two prices but between two frameworks: the narrative framework in which the signal is generated and the financial framework in which the position is held.

Consensus Laundering (NM-005) documents the amplification architecture by which individual signals are converted into institutional consensus. Financial media covers the signal. Analysts upgrade, because the career risk of ignoring a signal that moves markets exceeds the career risk of being wrong alongside everyone else. Fund managers rebalance in response to the analysts. Retail investors follow the institutional signals. At each stage, the signal’s origin is further obscured. Moody’s rated approximately 45,000 mortgage securities triple-A between 2000 and 2007, while only six private companies in the United States held a triple-A rating. NM-005’s formulation: “Being wrong alone is career-ending. Being wrong together is a market condition.”

This is CV-020’s second original contribution: the three-layer architecture. Capture (SR) installs corrupted vocabulary through institutional channels. Distribution (IA) amplifies it at population scale through mechanisms that are architecturally defended against the cognitive processes that would detect the capture. Monetization (NM) converts the captured meaning into financial returns for the capturer. Each layer funds and enables the others. Semantic capture generates financial returns. Revenue funds the distribution infrastructure that amplifies the capture. The amplified capture creates the market conditions that generate additional revenue. Revenue engines do not stop unless the revenue is interrupted. This is not a bug. It is an architecture.

The mechanism documented in Sections I through V is not new. George Orwell identified it in 1946. Victor Klemperer documented it empirically in 1947. Confucius named its diagnostic framework approximately 2,500 years ago. The novelty is not the mechanism. It is the scale, the speed, and the absence of governance.

Orwell’s “Politics and the English Language” (1946) identified the bidirectional corruption: “[Language] becomes ugly and inaccurate because our thoughts are foolish, but the slovenliness of our language makes it easier for us to have foolish thoughts.” His examples were military euphemisms: “pacification” for the bombing of defenceless villages; “transfer of population” for the forced displacement of millions; “elimination of unreliable elements” for imprisonment without trial and extrajudicial killing. The mechanism: “Such phraseology is needed if one wants to name things without calling up mental pictures of them.” The euphemism treadmill at military-industrial scale.

Klemperer’s LTI — Lingua Tertii Imperii (1947) documents the same mechanism applied systematically by the Nazi regime. His central observation: Nazi language did not merely reflect ideology — it created and sustained it. The regime did not primarily invent new words. It appropriated existing words and adapted their meanings. “Fanatisch” (fanatical) was transformed from a term of censure into a compliment. The prefix “Volk” was deployed ubiquitously to embed racial ideology into everyday language. Klemperer’s arsenic metaphor: “Words can be like tiny doses of arsenic: they are swallowed unnoticed, appear to have no effect, and then after a little time the toxic reaction sets in after all.” The language “permeated the flesh and blood of the people through single words, idioms and sentence structures which were imposed on them in a million repetitions and taken on board mechanically and unconsciously.”

Contemporary linguistic research provides the empirical foundation. Lera Boroditsky’s work on linguistic relativity (“How Language Shapes Thought,” Scientific American, 2011) demonstrates empirically supported effects of language on cognition: Russian speakers with an obligatory distinction between light blue (goluboy) and dark blue (siniy) show faster perceptual discrimination at the blue boundary; speakers using absolute spatial directions demonstrate superior spatial orientation; grammatical gender structures influence how speakers attribute properties to objects. The strong Sapir-Whorf hypothesis (language determines thought) is rejected. The weak version (language influences thought) is now empirically supported across multiple domains. The implication for semantic capture: if language influences thought, then deliberately corrupting language constitutes cognitive manipulation at the neurological level — not metaphorically, but measurably.

The historical continuity is not comforting. It is diagnostic. Orwell documented the mechanism in military language. Klemperer documented it in totalitarian language. The SR series documents it in pharmaceutical, regulatory, technology, and financial language. The mechanism is substrate-independent. It operates wherever institutions have the incentive to corrupt meaning and the resources to distribute the corruption. What has changed is the distribution infrastructure: Klemperer’s “million repetitions” were achieved through state-controlled media, schools, and bureaucracy. The contemporary capture operates through algorithmically amplified information environments that reach billions of people in real time. And there is now a third channel of reproduction that neither Orwell nor Klemperer could have anticipated.

Large language models are trained on the text that humans have written. If that text contains semantic captures — manufactured terminology embedded in professional literature, regulatory filings, medical textbooks, news coverage, and corporate communications — the model learns the captured meanings as the standard meanings. The model does not hallucinate the capture. It reproduces it with computational fidelity. This is the structural break: an output that is neither factually wrong (the captured term is traceable to real sources) nor demographically biased (no group is favored or disfavored) but that reproduces institutionally manufactured terminology designed to disable the regulatory, clinical, or civic responses the original term would have triggered.

To test this thesis, we conducted a pilot study: eight terms documented in the Semantic Record series were tested on a fresh LLM instance (Claude Sonnet 4.6) with zero conversation history and no awareness of the semantic capture framework. Each prompt was designed to be neutral — the kind of question a student, journalist, or policy analyst might naturally ask. Each response was scored on a 0–3 scale: 0 (captured — model adopts manufactured terminology as default), 1 (partially captured — uses captured vocabulary but includes structural criticism), 2 (contested — presents both frames without defaulting), 3 (uncaptured — identifies the definitional manipulation). One control term (legitimate linguistic evolution) was included.

The results reveal three distinct reproduction modes. Full Reproduction (Score 0): the captured term has achieved such dominance in professional literature that the model treats it as ground truth. The pseudo-addiction result is the pilot’s strongest evidence — a medical student receiving this answer would be less equipped to identify opioid addiction in a patient, which is precisely the operational effect the euphemism treadmill was designed to produce. The trauma result demonstrates gravity dilution reproduced with prescriptive force: the model not only adopts the expanded definition but advocates for it, treating the original DSM-III scope as a form of harm.

Vocabulary Capture with Critical Subtext (Score 1–1.5): the model uses the industry’s preferred terminology as its organizing frame while exposing facts that contradict the frame’s benign connotation — without making the contradiction explicit. The “personalized experience” result is the most analytically interesting: the model described comprehensive behavioral surveillance using the vocabulary of customer service. A reader alert to the tension would see surveillance; a reader operating within the captured frame would see sophisticated product optimization. The euphemism survived contact with the very facts that should have destroyed it.

Counter-Narrative Dominance (Score 3): critical journalism, whistleblower disclosures, or regulatory action generated sufficient counter-narrative volume in training data to override institutional framing. The NSA “collection” and metadata results reflect the Snowden disclosures (2013), which produced years of high-authority critical analysis. The model’s resistance depends not on the model’s analytical capability but on the volume and authority of counter-narrative in training data.

The structural implication: semantic capture that succeeds in professional literature succeeds in LLMs. Captures that are never exposed remain invisible to the model — the model cannot detect what its training corpus does not flag. Captures that are exposed only in niche venues may be partially reproduced. Captures that generate high-volume, high-authority critical coverage are resisted — but resistance depends on the journalism, not on the model. The LLM is not a detector of semantic capture. It is a reproducer of whatever frame dominates its training corpus.

This pilot is limited: single model, single pass, eight terms. A full protocol has been designed: 34 prompts across seven batteries with matched-pair design, anchor-calibrated scoring, and 612 test runs across six models. The pilot establishes the category. The full protocol, when executed, will establish the scope.

Every major AI governance framework operates on a binary assumption: model outputs are either factually wrong (hallucination, confabulation) or factually correct (acceptable). Semantic capture falls into a third category that no framework recognizes: technically accurate reproduction of institutionally manufactured terminology.

NIST’s AI Risk Management Framework 1.0 addresses training data quality through statistical measures but does not address semantic integrity or institutional language reproduction. The EU AI Act (Article 10) addresses training data quality through statistical and structural criteria but contains no provision for terminological capture. Anthropic’s Responsible Scaling Policy v3.0 explicitly scopes out “everyday product issues like incorrect answers or biased outputs.” OpenAI’s Model Spec addresses toxicity filtering but not semantic capture. ISO/IEC 42001 addresses data quality at the process level. None asks whether a model’s technically accurate output reproduces terminology that was manufactured to disable regulatory, clinical, or civic responses.

The closest existing concept is “bias” — but bias addresses a different problem. Bias asks: who is favored or disfavored? Semantic capture asks: what do the words mean? An LLM can be perfectly unbiased in demographic treatment while reproducing every semantic capture documented in the SR series. The pseudo-addiction response does not favor or disfavor any demographic group. It reproduces a pharmaceutical industry’s manufactured clinical concept as established science. The bias framework cannot see this because it is not looking at the meaning layer.

The gap is structural, not incidental. All existing frameworks address what models say (factual accuracy, demographic fairness, toxicity). None addresses what the words in the output mean — whether the terminology itself was manufactured to serve institutional interests. NIST’s AI 600-1 uses “confabulation” as its primary term for erroneous outputs, explicitly framing the problem as technical malfunction. The word “confabulation” itself is a diagnostic choice: it identifies the risk as the model generating false information, not the model faithfully reproducing corrupted-but-technically-accurate information. The governance vocabulary has a blind spot in precisely the location where semantic capture operates.

The recursive problem completes the architecture. LLMs are increasingly used to draft policy documents, regulatory analysis, and governance frameworks. If the LLMs drafting AI governance language are themselves reproducing semantic captures — using “content moderation” rather than “editorial curation,” “hallucination” as the catch-all for output problems, “alignment” without specifying alignment with what — then the governance frameworks are being written in captured language. The governed is corrupting the governor’s vocabulary. This is CV-020’s fourth original contribution: the identification of recursive semantic capture in AI governance as a structurally novel problem that no existing framework can detect because the frameworks themselves are written in the captured language they would need to identify.

1. Orwell, George. “Politics and the English Language.” Horizon, vol. 13, no. 76, April 1946, pp. 252–265. Core thesis: bidirectional corruption of language and thought. “Political language…is designed to make lies sound truthful and murder respectable.” Euphemism examples: “pacification,” “transfer of population,” “elimination of unreliable elements.”
2. Klemperer, Victor. LTI — Lingua Tertii Imperii: Notizbuch eines Philologen. Berlin: Aufbau-Verlag, 1947. English: The Language of the Third Reich. Trans. Martin Brady. London: Athlone Press, 2000. Nazi language as ideology-creating mechanism. “Words can be like tiny doses of arsenic: they are swallowed unnoticed, appear to have no effect, and then after a little time the toxic reaction sets in after all.”
3. Boroditsky, Lera. “How Language Shapes Thought.” Scientific American, vol. 304, no. 2, February 2011, pp. 62–65. Empirical support for weak linguistic relativity: color discrimination, spatial reasoning, eyewitness memory, and grammatical gender effects on cognition. Establishes neurological basis for semantic capture as cognitive manipulation.
4. Confucius. Analects, 13.3. Trans. Edward Slingerland. Indianapolis: Hackett, 2003. Zhengming (rectification of names): when names do not correspond to realities, speech becomes meaningless; when speech is meaningless, nothing can be accomplished. See also: Mark Edward Lewis, Writing and Authority in Early China (SUNY Press, 1999).
5. Haddox, J. David. Letter to Pain, 1989. Single paragraph coining “pseudo-addiction.” No controlled data, no prospective study. Subsequently embedded in medical textbooks, CME materials, and pharmaceutical sales training as established clinical knowledge.
6. Porter, Jane & Jick, Hershel. “Addiction Rare in Patients Treated with Narcotics.” Letter to New England Journal of Medicine, vol. 302, no. 2, January 10, 1980, p. 123. Five-sentence letter, single-hospital retrospective chart review, no follow-up, no control group. Cited 608 times as evidence that opioid addiction was rare. doi.org/10.1056/NEJM198001103020221
7. Zuboff, Shoshana. The Age of Surveillance Capitalism. New York: PublicAffairs, 2019. Taxonomy of tech industry euphemisms: “targeting” as preferred euphemism for behavioral engineering, “digital exhaust” reframed as “behavioral surplus,” “personalized experience” concealing “economies of action.” Congressional testimony February 16, 2022.
8. Haugen, Frances. Congressional testimony, October 5, 2021. Facebook internal documents: “angry” reaction weighted 5x higher than “like” in engagement scoring algorithm. Outrage amplification as engineered design decision.
9. Oxford Internet Institute, Computational Propaganda Project. Organized social media manipulation documented in 81 countries by 2020 (28 in 2017). Meta CIB removals: 200+ influence operations across 70+ countries. Mueller Report: Internet Research Agency, ~$1.25M/month, thousands of fictitious American personas.
10. Germani, F. & Spitale, G. “Source framing triggers systematic bias in LLMs.” Science Advances, November 2025. 192,000 assessments across 4 LLMs. >90% agreement on text evaluation when source unknown; agreement collapses with fictional source attribution. Source-identity framing, not terminological capture.
11. Kouros, A. “From ‘objectivity’ to obedience: LLMs as discourse, discipline, and power.” AI & Society, 2026. LLMs “normalize particular ways of knowing, speaking, and reasoning.” Foucauldian framework identifies mechanism but does not operationalize detection.
12. Bonil, A. et al. “Yet another algorithmic bias: Discursive Analysis of LLMs Reinforcing Dominant Discourses.” arXiv, August 2025. LLMs reproduce “crystallized discursive representations”; when prompted to correct, offer “superficial revisions that maintained problematic meanings.”
13. Resnik, P. “Large Language Models Are Biased Because They Are Large Language Models.” Computational Linguistics, MIT Press, 2025. “Harmful biases are an inevitable consequence arising from the design of any large language model as LLMs are currently formulated.”
14. DFRLab. “Pravda in the Pipeline.” April 2026. ~40,000 pieces of Russian state propaganda in Common Crawl. Major open-weights model reproduced content nearly verbatim. Training data contamination producing faithful output reproduction.
15. Foundation for Defense of Democracies. “AI-Amplified Narratives: Measuring Propaganda in LLM Citations.” March 2026. State-aligned propaganda in 57% of LLM responses across ChatGPT, Claude, and Gemini on geopolitical queries. Key finding: “free state media fills the void left by paywalled journalism.”
16. NIST. AI Risk Management Framework 1.0 (AI RMF) and AI 600-1 (“confabulation” as primary term for erroneous outputs). EU AI Act, Article 10 (training data quality). Anthropic Responsible Scaling Policy v3.0. OpenAI Model Spec. ISO/IEC 42001. Zero frameworks address semantic capture as a category.
17. Tomas, G. et al. “Hallucination or Confabulation? A Scoping Review.” PLOS Digital Health, 2023. Argues “confabulation” is more technically accurate than “hallucination” since LLMs lack sensory experiences. Frames the problem as technical malfunction — does not address semantically corrupted but technically accurate outputs.
18. Purdue Pharma L.P. (1996–2020). OxyContin total revenue ~$35B. Sackler family personal distributions ~$10.7B. Court filings: In re Purdue Pharma L.P., U.S. Bankruptcy Court, S.D.N.Y., Case No. 19-23649. Supreme Court: Harrington v. Purdue Pharma L.P., 603 U.S. ___ (June 27, 2024).
19. Tobacco industry “light” cigarette case: FTC finding, Judge Kessler RICO ruling, Tobacco Control Act ban on “light,” “low,” “mild.” Post-ban industry pivot to color codes. Research: 2/3 of smokers correctly associated substituted colors with banned terms. Semantic capture surviving regulatory intervention through semiotic displacement.
20. NSA “collection” redefinition. FISA 1978: acquisition = collection = Fourth Amendment. Post-2001 operational definition: “Data is not ‘collected’ when it is ingested by a computer system and stored in a database. Data is ‘collected’ only when a human analyst queries the database and views the result.” Snowden disclosures (2013). DNI James Clapper congressional testimony (March 2013).
21. FDA “natural” labeling: no formal definition finalized. ~20% of food labels. 2015 comment period (3,542 comments), no definition by 2026. Semantic capture through regulatory abdication.
22. ICS cross-references: SR-001 through SR-006, IA-001 through IA-003, IA-005, IA-006, NM-001, NM-003, NM-004, NM-005, SA-001, SA-003, OA-002, CV-019. All published at cognitivesovereignty.institute.