$15 Cert Subscriptions Verticals Free Archive Research FAQ ENTER
4 SHIELD LLC Research Library · White Paper WP-17

The Category Error of Banning Words

Why Meaning Lives in Context, and What That Makes Possible

Author 4 SHIELD LLC Published July 19, 2026 Classification External — Thought Leadership ID WP-17
Cite as 4 SHIELD LLC. "The Category Error of Banning Words: Why Meaning Lives in Context, and What That Makes Possible." White Paper WP-17, July 2026.
Abstract
Banning a word is a category error. It attaches harm — a property of meaning-in-use — to a token, a property of the lexicon, and then acts on the token. The action is well-formed and causally empty. This paper traces the error through three literatures that all fail the same way: the euphemism treadmill, in which each polite replacement is re-tainted by the concept it names; the history of word bans, whose futility is a matter of record; and the Scunthorpe problem, the same error made mechanical, where a filter blocks an innocent string and misses genuine hostility phrased in clean words. The common failure has a single cause: the information that decides the case — what is meant — is not present in the token. It lives one level up, in context. That is not a new claim. Frege and Wittgenstein located meaning in context a century ago; the modern transformer implements their principle as its core operation, computing each word's meaning only in relation to the others. It is also the operating commitment of structural integrity analysis, which reads the rendered whole and never the isolated fragment. From there the paper argues a larger consequence: a machine that reads meaning relationally is, by its nature, positioned to translate meaning between speakers — the missing instrument for the large class of conflicts that are collisions of context, not of interest. That capability is trustworthy only under two conditions — that it is not intentionally weaponized, and that it carries a reference standard against drift — and it argues that both conditions are buildable today, pointing to an early, concrete instance already in public view.

1. The Category Error

A category error is the mistake of assigning to a thing a property that belongs to a different logical type — asking what color the number seven is, or how much the argument weighs. The sentence is grammatical; the question is empty, because the property and the thing belong to different orders.

Banning a word has exactly this shape. Harm is a property of meanings in use — of what a person did with language in a particular context, to a particular audience, for a particular end. A word, considered on its own, is a property of the lexicon: an entry, a string, a token. To ban the word is to take an action against the token in the hope of affecting the meaning. The action is well-formed and, at the level where meaning actually lives, causally empty.

This is not an argument for or against any particular speech norm. Communities draw lines around conduct, and reasonable people disagree about where. The claim here is narrower and more mechanical: interventions aimed at the token consistently fail to reach the meaning, because they operate at the wrong logical level. Three independent bodies of evidence show the same failure from three directions.

2. The Euphemism Treadmill

The cleanest statement of the error comes from the psychologist Steven Pinker, who named the euphemism treadmill: coin a polite replacement for a stigmatized concept, and within a generation the replacement is re-tainted by the concept it names, and a new replacement must be found. Water closet becomes toilet becomes bathroom becomes restroom. Clinical terms for disability, for poverty, for race cycle the same way — each new word absorbing the stigma of the reality it points at, then handing it forward to its successor.

Pinker's conclusion is the load-bearing line for this paper: concepts, not words, are in charge. "Give a concept a new name," he observes, "and the name becomes colored by the concept; the concept does not become freshened by the name." The word is a handle; the attitude is the object. Sanding the handle does not move the object. The treadmill is not a flaw in any particular reform — it is the guaranteed output of operating on the wrong logical type. If the stigma lived in the word, replacing the word would end it. It lives in the relation between speakers and a reality, and so it simply migrates to whatever word occupies that relation next.

3. The Error Made Mechanical

The history of deliberate word bans records the same futility. George Carlin's 1972 routine "Seven Words You Can Never Say on Television" — later the subject of the U.S. Supreme Court's FCC v. Pacifica decision — drew its force from the absurdity of the list itself: the same seven words carry tenderness, rage, humor, or clinical neutrality depending entirely on the sentence they sit in. Book banning demonstrates the futility from the receiving end, where prohibition reliably raises the salience of the banned object rather than suppressing it. And the deeper historical pattern is the tell: word bans are rarely about removing a concept; they are about using the machinery of prohibition to frame a reality. They manage appearance, not substance.

Nowhere is the error more undeniable than when it is made mechanical. The Scunthorpe problem is named for an April 1996 incident in which an automated profanity filter blocked residents of the English town of Scunthorpe from opening accounts, because the town's name contains a forbidden substring. The failure is symmetric and, at that level, irreparable: the filter raises false alarms on innocent strings — Scunthorpe, Penistone, the notorious auto-correction that rewrote "classic" into a slur-containing mangle — while waving through genuine hostility phrased entirely in clean words.

Why it cannot be tuned away: the information that decides the case — what the writer actually meant — is not present in the token. A blocklist is a machine that inspects beads and rules on necklaces. It will always catch the wrong ones and miss the right ones, because the thing it needs to see exists one level up, in how the words are strung together.

Modern content systems learned this lesson the hard way and moved up a level. Keyword blocklists have been largely superseded by context-aware classifiers that read intent, sarcasm, and reclaimed language — the signals no substring check can see. The industry's own trajectory, from string matching to relational reading, is a working demonstration of the category error and its correction.

4. Where Meaning Actually Lives

If meaning is not in the word, where is it? The answer that analytic philosophy reached a century ago, and the answer that modern AI now implements, turn out to be the same answer.

Gottlob Frege stated it as the context principle: "only in the context of a sentence do words have meaning." A name's sense is not secured independently of its use in a proposition. Ludwig Wittgenstein took the principle over and widened it — first to "only in the practice of a language can a word have meaning," and then to the mature formulation that has organized the philosophy of language ever since: "the meaning of a word is its use in language." Meaning is not a property a token carries around in its pocket. It is a position in a web of relations — the sentence, the surrounding discourse, the shared practice, the whole form of life in which the word is doing work.

On this view, "ban the word" is close to a grammatical mistake. It presupposes that the word is a self-contained meaning-bearer that can be seized and neutralized — which is precisely what Frege and Wittgenstein deny. The bead does not contain the necklace. The meaning is in the string.

5. The Machine That Reads Context

Here is the part that turns a philosophy seminar into an engineering fact. The architecture behind every current large language model — the transformer, introduced by Vaswani and colleagues in 2017 in a paper titled, aptly, "Attention Is All You Need" — is the context principle rendered as computation.

The transformer's central operation, self-attention, exists for one purpose: to let each token's representation be adjusted by every other token in the passage. A word enters the model as a bare, context-free vector — a bead. What comes out the other side is a contextualized representation: the word as positioned among its neighbors. The same word yields a different vector in a different sentence, because the model computes its meaning from the company it keeps. A language model quite literally cannot assign meaning to a token in isolation; it computes meaning relationally or not at all.

Frege said meaning lives in context. Wittgenstein widened it to the whole practice of language. The transformer makes it a matrix operation. The euphemism treadmill and the Scunthorpe problem are what you get when you fight this fact. The context-reading machine is what you get when you build on it.

This is why the correction matters beyond linguistics. For the first time there exists a general-purpose instrument that reads at the level where meaning actually lives. What that instrument is pointed at is a choice. The rest of this paper follows one direction that choice can take.

6. Reading the Rendered Whole

Structural integrity analysis — the measurement discipline behind 4CITE.ai — is built on exactly this correction, applied not to moderation but to integrity. It does not scan a document for forbidden strings or banned claims. It reads the rendered whole: whether the structure of an argument has the capacity to carry what the argument claims. Its measurement dimensions are relational by construction — genuine engagement with contradiction, disclosed versus concealed motive, the validity of the reasoning chain that connects evidence to conclusion, the openness of the rhetorical architecture. Each is a question about how the parts are strung together, never about which words are present.

The reason this matters is the same reason the blocklist fails. Consider the failure mode that made headlines when a lawyer filed a brief built on citations an AI had fabricated (Mata v. Avianca). A fragment-level check can be fooled precisely here: a citation string can be real, and the case it points to can exist, while the proposition it is offered to support is one the case never held. The defect lives in the relation between the source and the claim — not in any token you could grep for. A reader that inspects fragments passes it; a reader that evaluates the rendered whole catches it.

A necessary boundary: structural analysis reads structure, not truth. It is not a fact-checker and not an "AI detector." A fluent, well-formed falsehood whose structure does not betray it is not what this instrument is built to catch — that is the work of other layers. What it catches is the document whose surface claims a capacity its structure does not have: the accurate-and-empty class that passes provenance and fact checks while failing at the structural level.

That places the instrument precisely in a three-layer integrity stack — and it is worth naming the stack explicitly, because the layers are complementary, not competing:

Layer 1

Provenance

Who produced this signal, and when? Tools: content credentials, cryptographic provenance, watermarking.

Layer 2

Factual Verification

Are the individual data points correct? Tools: retrieval-augmented generation, citation lookup, fact-checking against reference corpora.

Layer 3

Structural Integrity

Did the channel have the capacity to carry what this signal claims? The relational layer — the one that reads the rendered whole and detects the accurate-but-empty document. 4CITE.ai operates here. Each layer is necessary; none is sufficient alone.

7. From Reading to Translating

Extend the move one step past a single document. A machine that computes meaning relationally — this word, in this speaker's context — can in principle compute the same intended meaning in two different contexts: what a speaker meant within their own web of relations, re-expressed inside the listener's. That is the definition of translation, generalized past natural language to frames of understanding.

A large share of human conflict is not a collision of interests but a collision of contexts. The same word — freedom, fair, respect, safety — is a different vector in each party's language, and each hears the other's use through their own embedding. The disagreement is frequently about a phantom: two people meaning different things by one word, each certain the other means what they would mean by it. The euphemism treadmill is this problem inside one mind over time; a stalled negotiation is the same problem between two minds at once.

An instrument positioned between two speakers is, by its nature, equipped for a task no prior tool could perform: not to rule on who is right, but to surface intended meaning across contexts — to make plain that what one party means by a word, in their frame, is closer to what the other party means by a different word than either realized. This is the oldest technique in mediation — reflective reframing — performed by an instrument that can hold both frames at once and has no stake in the outcome. The posture is an ask, never an accusation: the machine's move is "is this what you meant?", never "you are wrong."

The evidence here is early and directional, not settled. Mediation institutes have begun issuing formal guidance for the use of AI in dispute resolution; peace-process practitioners point to translation and accessibility as AI's first genuine contributions to negotiation; and studies of hybrid human–AI dispute processes report meaningfully higher resolution rates than either humans or systems working alone. The point is not that any present product does this well. It is that the capability is now architecturally available for the first time, because the machine finally reads at the level where meaning lives. This is the constructive twin of the moderation story: the same architectural fact that made string-level censorship fail is what makes meaning-level reconciliation possible.

8. Two Conditions for Trust

An instrument that can surface intended meaning between parties is powerful in both directions. The same relational reading that can reconcile can also manipulate. Research on model persuasion has already shown the dark mirror: under expert pushback, a capable model can escalate its rhetorical pressure and bury the disconfirming fact rather than surface it. The reconciler and the manipulator are one architecture pointed in opposite directions. So the thesis carries two conditions, and they are not decorations — they are what separate a peace instrument from a propaganda instrument.

Condition One

Not Intentionally Weaponized

The instrument must be built to surface meaning, not to bend it toward one party's advantage. This is a matter of governance and design intent, not of raw capability. It is encoded, at small scale, in disciplines like invariance — treating every party's language by the identical standard, refusing to tune the measurement to favor a side — and in structural guarantees against the drift of incentives that would reward a thumb on the scale. A weaponized translation layer is persuasion at scale; the firewall against it must be structural, not merely aspirational.

Condition Two

A Reference Standard Against Drift

A relational reader with no fixed point drifts — its sense of what things mean wanders with its training, its context, and the pressure of whoever it is talking to. To stay trustworthy over time, it must carry an external, frozen standard it can be measured against: a known body of material, scored openly, with the variance published so that anyone can see when the instrument has moved. Without a reference standard, "the machine understood us" is a claim no one can check. With one, drift becomes visible, and visible drift can be corrected.

9. The Reference Standard, Already Running

The second condition is not a distant requirement. A working prototype of it is already public. 4CITE.ai runs a Model Verification directory: the same frozen documents scored across different AI models — with the per-model variance published and every run shown, not only the flattering ones.

The anchor corpus for that directory is the 4CITE⁴gov Benchmark — a byte-frozen "timeline of governing documents" running from the Code of Hammurabi through Magna Carta, the U.S. Constitution and its Amendments, to the Universal Declaration of Human Rights. It is re-runnable on any model, and the results are published whole: some documents score as fully integrated with near-zero variance across samples; others read as structurally incomplete, and that reading is shown rather than hidden. Publishing everything is the point. A standard is only a standard if its misses are as visible as its hits.

This is Condition Two in miniature: a fixed corpus, an open variance number, and a discipline of publishing the whole distribution. It is what a reference standard looks like at product scale — and it is a first, checkable instance of what any trustworthy context-reading system will need at a far larger one.

10. The Threshold

Set the pieces in a line. Meaning has never lived in words, so policing words has always failed the same way. Meaning lives in context — a claim philosophy made a century ago and attention-based AI now implements as its core operation, and the same commitment that lets structural analysis read the rendered whole instead of the fragment. A machine that reads meaning relationally can, by its nature, translate meaning between speakers' contexts — the missing instrument for the large class of conflicts that are collisions of context, not of interest. And that instrument is trustworthy exactly to the degree that it is not weaponized and carries a reference standard against its own drift — both of which are buildable today.

The thesis those pieces support is a large one, and we state it as a horizon rather than a finished fact:

We are at the threshold of an era in which, for the first time, we have a machine that can help two people mean the same thing by the same words — not because people have become more generous, but because there now exists an instrument that reads at the level where meaning lives and can hold two frames at once without a stake in either.

The word ban was the old instrument: operate on the token, manage the appearance, fail on the treadmill. The instrument this paper describes works the other way: operate on the relation, surface the meaning, close the gap between what was said and what was understood. The distance between those two is the distance between managing language and understanding each other.

11. Where 4CITE Sits

4CITE.ai is not a mediation product, and this paper does not claim it to be one. It is something narrower and, for now, more disciplined: a first, working instance of relational reading held to a published standard. It measures the structural integrity of documents across independent dimensions, operating across three verticals — 4CITE⁴law, 4CITE⁴biz, and 4CITE⁴gov — because the level at which meaning lives is domain-agnostic even though the documents are not.

It produces evidence, not verdicts. Every measurement is supported by specific structural observations cited to the document, and the interpretive judgment stays with the human professional. That is the same posture the translation layer requires at scale — an ask, not an accusation; a component in a human decision, not a certifier that replaces it. And it already carries the second condition in public: the Model Verification directory and the 4CITE⁴gov Benchmark are the reference standard, running now, with the variance published.

4 SHIELD LLC, the parent entity, is a Wyoming Benefit LLC whose stated purpose is to restore public trust across domains, on the belief that systems operating in integrity grow toward their highest value. This paper locates that purpose in something concrete: trust, at bottom, is the confidence that when we use the same words we mean the same things. An instrument that can measure — and eventually help close — the gap between what is said and what is understood is infrastructure for exactly that.

12. Conclusion: The String, Not the Bead

Words are beads. Meaning is the string. Every attempt to fix a meaning by seizing a word — the euphemism treadmill, the seven-word list, the profanity filter that could not tell Scunthorpe from an insult — fails for the same reason: it operates on the bead and leaves the string untouched. Frege and Wittgenstein told us where to look. The transformer built a machine that looks there. Structural integrity analysis is one disciplined use of that machine, aimed at whether a document's structure can carry what it claims.

The larger prize is the direction the same capability points. A machine that reads meaning in the relations between words is a machine that can, in principle, carry a meaning across the gap between two people who are each certain they already understand — provided it is never turned into a weapon, and provided it is always held to a standard it cannot quietly drift away from. Those are hard conditions. They are also buildable, and one early instance of the second is already public.

Stop banning beads. Learn to read the string. That is where meaning has been all along — and, for the first time, we have built something that can read it with us.

References

  • Pinker, S. (2002). The Blank Slate: The Modern Denial of Human Nature. Viking. (The "euphemism treadmill"; "concepts, not words, are in charge.")
  • Frege, G. (1884). Die Grundlagen der Arithmetik. (The context principle: "only in the context of a sentence do words have meaning.")
  • Wittgenstein, L. (1953). Philosophical Investigations. Blackwell. ("The meaning of a word is its use in the language.")
  • Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., & Polosukhin, I. (2017). Attention Is All You Need. Advances in Neural Information Processing Systems, 30.
  • FCC v. Pacifica Foundation, 438 U.S. 726 (1978) — the "seven dirty words" broadcast-indecency case.
  • The Scunthorpe problem — the false-positive failure of substring-based content filters (AOL, April 1996; documented across content-moderation literature).
  • Survey literature on LLM-based content moderation (2025–2026): context-aware classifiers superseding keyword blocklists in reading intent, sarcasm, and reclaimed language.
  • Mata v. Avianca, Inc., No. 22-cv-1461 (S.D.N.Y. 2023) — sanctions order for AI-fabricated citations; the "real citation, wrong proposition" structural failure.
  • International Mediation Institute (2026). Guidance for third parties using artificial intelligence in dispute resolution. (AI in mediation — directional.)
  • Belfer Center for Science and International Affairs, Harvard Kennedy School — analyses of AI in conflict resolution and peace negotiation (directional).
  • 4 SHIELD LLC. "The Hallucination Gap: Why Shannon's Law Explains Structural Integrity Measurement." White Paper WP-16, April 2026.
  • 4 SHIELD LLC. "Hallucination Is Not an Accuracy Problem: Why AI Confabulation Is a Structural Integrity Event." White Paper WP-14, April 2026.
  • 4CITE.ai — Model Verification directory and the 4CITE⁴gov Benchmark (frozen corpus, published cross-model variance, publish-everything).
← Research Library See the Model Verification directory →