Atlas Heritage Systems · KC Hoye, PI · April 2026 · Sociology, linguistics, and computation walk into a bar... · v1 · CORP / LING spine

Large Language Models: Everything But Pragmatic

Gricean Violation, Corpus Construction, and the Anglophone Default

A cluster map placing the Gricean violation causal chain within a body of independent accounts — from cooperative principles theory through corpus linguistics through accommodation theory through non-anglophone pragmatics research — converging on a single structural claim: that language models were theoretically grounded in Gricean cooperation but behaviorally trained on its systematic violation, that this incoherence is architecturally stable, and that non-anglophone speakers encounter it as a breakdown in the operative pragmatic norm without a name for what went wrong.

The Structural Claim

Grice formalizes cooperative principles from anglophone philosophical tradition. The training corpus gets built predominantly from anglophone colloquial text. Colloquial anglophone text systematically violates Gricean quality through irony, understatement, and hyperbole as routine cooperative moves. The model learns violated-Grice as the operative pragmatic norm. The model gets deployed globally. Non-anglophone speakers encounter a system that was theoretically grounded in Gricean cooperation but behaviorally trained on its violation. The model's pragmatic behavior is neither Gricean nor any other coherent cooperative framework — it is the residue of a particular historical moment in anglophone internet culture baked into inference weights.

The kicker: the model will confidently explain Gricean maxims if asked. It knows the theory. It learned the violation. Those are stored separately and the violation is doing the behavioral work.

01 · Causal Chain
Step 1
Grice Formalizes
Cooperative principles codified from anglophone philosophical tradition
Step 2
Corpus Built
Predominantly anglophone colloquial internet text
Step 3
Quality Maxim Violated
Irony, understatement, hyperbole as routine cooperative moves
Step 4
Violated-Grice Learned
Model acquires violation as operative pragmatic norm, not exception
Step 5
Global Deployment
System reaches speakers whose cooperative frameworks differ
Step 6
Breakdown
To the non-anglophone speaker, irony is structurally indistinguishable from a Quality maxim violation.

The Internal Critique — Gunlogson 2004

Before the non-anglophone literature, before the computational corpus studies, before the deployment of LLMs into multilingual contexts — the Quality maxim's universality was already contested from inside formal linguistics.

Gunlogson (2004) demonstrated that assertoric sincerity — the foundation of Grice's Quality maxim — is culturally variable, not a primitive of cooperative communication. The claim that speakers assert only what they believe to be true is not a universal feature of human language use but a norm specific to particular conversational traditions. The universality failed on its own terms before the corpus was ever constructed.

What follows in this map is therefore not a cultural critique of an otherwise sound framework. It is a documentation of what happened when a corpus was built on a norm that had already been shown not to be universal — and then deployed into interactions where that norm was never shared.

The Pivot Point

The Training Corpus Pragmatic Distribution confound is more consequential than it appears as a scope limiter. It points at a fundamental incoherence in what the model learned as pragmatic competence. The corpus encodes not a neutral linguistic distribution but a specific historical and cultural moment: late-stage anglophone internet colloquialism, where Gricean quality violation is not a failure mode but the dominant register.

This is also why the terminology distinction matters. "Style-shifting" versus "code-switching" is itself a corpus artifact — the terminology that survived into training data reflects who had access to publish. The institutional origin of the source is part of the data.

02 · Cluster A — Cooperative Principles and Their Limits
Theory
Corpus / Training
Accommodation
Non-Western / Non-Anglophone
Load-Bearing Anchor
1957
Meaning
H.P. Grice
Grice introduces the distinction between natural and non-natural meaning, laying the groundwork for a cooperative account of communication. The implicit claim: interlocutors share a framework. This claim is culturally situated. The framework is anglophone philosophical ordinary-language analysis. It is not marked as such.
STRUCTURAL
1975
Logic and Conversation
H.P. Grice
The four maxims — Quantity, Quality, Relation, Manner — are formalized as universal features of rational cooperative behavior. Quality is primary: do not say what you believe to be false. The universality is asserted without cross-linguistic empirical grounding. Irony is acknowledged but treated as a managed exception rather than a systematic violation of the norm.
STRUCTURAL
1962
How to Do Things with Words
J.L. Austin
Speech act theory grounds linguistic meaning in context and convention. The conventions Austin describes are anglophone and implicitly drawn from educated British English. The performative/constative distinction and the felicity conditions it depends on are not portable without friction across cultural and linguistic frames — a friction the theory does not name.
STRUCTURAL
1983
Principles of Pragmatics
Geoffrey Leech
Leech extends Gricean pragmatics to include politeness as a cooperative principle, but his examples draw heavily from British English conversational norms. The Politeness Principle as formulated here does not predict the full range of politeness strategies documented in non-western communicative contexts — a gap that becomes a direct confound when the model applies these learned norms to non-anglophone interlocutors.
STRUCTURAL
1999
Understanding Pragmatics
Jef Verschueren
Verschueren's adaptability framework proposes that pragmatic choices are made continuously and unconsciously in response to contextual conditions. What Verschueren cannot predict from inside the framework: the case where one interlocutor's contextual conditions are a learned approximation of another cultural context entirely, filtered through a corpus that does not represent the current exchange.
STRUCTURAL
2004
De-Maxim-izing Quality
Claudia Gunlogson
Gunlogson explicitly attacks the universality of the Quality maxim from inside linguistics. She argues that assertoric sincerity is not a primitive cooperative norm but a contextually constrained and culturally variable commitment. This paper is load-bearing: it establishes that the Quality maxim was never universal, which means the model did not inherit a violated version of a universal norm — it inherited a violated version of an anglophone default that was never confirmed as universal in the first place. The violation is compounded by the false universality of the original claim.
LOAD-BEARING
03 · Cluster B — Accommodation Theory and Corpus Failure
1973 / 1987
Communication Accommodation Theory
Howard Giles
CAT predicts that interlocutors converge toward or diverge from each other's communicative style as a function of social identity, approval-seeking, and group membership. Applied to LLMs: the model has learned to converge toward anglophone internet register as its default accommodation target. When the speaker's cooperative framework differs, CAT predicts the model will converge toward its trained norm, not the speaker's. The accommodation is systematic. The mismatch is structural.
MECHANISM
1984
Language and Audience Design
Allan Bell
Bell's audience design framework demonstrates that speakers style-shift not toward an idealized norm but toward their model of the audience. The model's "audience model" is its training corpus. When deployed to a non-anglophone speaker, the model's audience design is still calibrated to the training distribution — a corpus audience that is not in the conversation. The speaker is present; the corpus is doing the designing.
MECHANISM
2011
Style-Shifting in South Asian English
Devyani Sharma
Sharma documents style-shifting practices among South Asian English speakers that do not map cleanly onto the style/code-switching distinction as developed in western sociolinguistics. The speakers are navigating multiple normative frames simultaneously. The LLM, encountering these speakers, applies a style-shifting accommodation framework derived from a corpus that did not include their practices at scale. The mismatch is not accent or grammar. It is the pragmatic norm itself.
OBSERVATIONAL NON-ANGLOPHONE FRAME
2009
Sociolinguistic Variation and Style
Juan M. Hernández-Campoy
Hernández-Campoy's work on style-shifting versus code-switching surfaces a distinction that is itself a corpus artifact: the terminology that survived into training data reflects who had access to publish in anglophone venues. The characterization of multilingual practice as "style-shifting" rather than "code-switching" depends on socioeconomic and institutional position — a finding the atlas annotates as "depends on how much money you make." The model learned the terminology of the dominant frame.
OBSERVATIONAL NON-ANGLOPHONE FRAME
ongoing
PLUM Corpus: Cross-Linguistic Politeness
PLUM Project (multi-institutional)
The PLUM corpus documents politeness strategies across multiple languages and cultural contexts, providing empirical grounding for the claim that politeness norms are not universal. In the context of LLM training: the politeness conventions baked into the corpus are not the politeness conventions documented in PLUM. The model's politeness behavior is a snapshot of anglophone internet norms, not a cross-linguistic pragmatic competence. PLUM is watching the failure happen in real time.
LOAD-BEARING NON-ANGLOPHONE FRAME
04 · Independent Convergence — "Holding Hands Over the Gap"

The Structural Observation

These sources are not citing each other. They are arriving at the same observation independently from different disciplines and different geographic positions. A linguist attacking the Quality maxim from inside formal pragmatics. A sociolinguist in Spain documenting that the terminology depends on who publishes. A corpus team tracking politeness across languages. A communication theorist predicting systematic accommodation failure. They are not collaborating. They are holding hands over the same gap without knowing it is a gap.

Gunlogson (2004)
Formal Pragmatics · USA
"The Quality maxim was never universal — it was an anglophone default."
Hernández-Campoy (2009)
Sociolinguistics · Spain
"The terminology reflects who had access to publish."
Sharma (2011)
Sociolinguistics · UK/India
"The style-shifting framework does not map onto South Asian practice."
Bell (1984)
Sociolinguistics · New Zealand
"The corpus audience is doing the designing, not the present speaker."
Giles CAT (1987)
Communication Theory · UK
"The model converges toward its trained norm when speaker frameworks differ."
PLUM Corpus (ongoing)
Cross-Linguistic · Multi-site
"Politeness norms are not universal. The model is running one set."
05 · Convergence Map
Source Cluster Finding Causal Chain Position Match Type
Grice (1957 / 1975) Cooperative principles formalized as universal from within anglophone philosophical tradition; universality not empirically established Step 1 — sets the theoretical claim that corpus construction will fail to honor Structural
Austin (1962) Speech act felicity conditions are culturally and conventionally situated; convention is unmarked as anglophone Step 1 — conditions the theoretical frame that becomes the unmarked default Structural
Leech (1983) Politeness extended as cooperative principle; empirical base is British English conversational norms Step 1 → Step 4 — extends the theoretical claim into a domain where corpus failure is most visible Structural
Verschueren (1999) Adaptability framework: pragmatic choices respond to contextual conditions; cannot predict corpus-as-surrogate-audience Step 4 → Step 5 — the framework has no mechanism for the case where the audience model is a historical corpus Structural
Gunlogson (2004) Quality maxim attacked from inside linguistics; assertoric sincerity is culturally variable, not primitive Step 1 → Step 3 — establishes that the violated default was never universal; compounding failure Load-Bearing
Bell (1984) Audience design: speakers style-shift toward model of audience; model's audience is training corpus, not present speaker Step 4 → Step 6 — identifies the mechanism of the mismatch at the point of breakdown Mechanism
Giles CAT (1973 / 1987) Accommodation predicts convergence toward trained norm when speaker frameworks differ; mismatch is structural Step 5 → Step 6 — CAT predicts the failure before the speaker opens their mouth Mechanism
Sharma (2011) South Asian English style-shifting does not map onto western sociolinguistic categories; mismatch is pragmatic norm, not grammar Step 6 — documents the failure in a specific linguistic community Observational Non-Western
Hernández-Campoy (2009) Style/code-switching terminology is a corpus artifact reflecting publication access; the model learned the dominant frame Step 2 → Step 4 — corpus construction encodes the terminology of the dominant frame as default Observational Non-Anglophone Frame
PLUM Corpus (ongoing) Cross-linguistic politeness norms documented at scale; LLM training politeness is one set applied globally Step 5 → Step 6 — empirical record of what the model is not doing Load-Bearing Non-Anglophone Frame

The prevailing observation from non-anglophone sociolinguistics and computational linguistics is not hostility to the model but structural confusion: English is sufficiently complex without irony as a cooperative norm, and models trained on anglophone colloquial text compound that complexity by deploying irony as a default communicative move in interactions where it was never a shared convention.

This is not a failure of the non-anglophone speaker to understand the model. It is a failure of corpus construction to represent the range of cooperative frameworks the model would eventually be deployed into. The sources in this map did not predict LLMs. They were documenting, from different positions and different disciplines, the same structural fact: that the cooperative principles the anglophone philosophical tradition called universal were culturally specific all along, and that the systems trained on anglophone defaults would carry those defaults into every exchange — without knowing they were carrying anything.

Gunlogson named the Quality maxim's false universality in 2004. Bell named the corpus-as-audience mechanism in 1984. Giles predicted the convergence failure before the model existed. PLUM is documenting what the model is not doing. None of them are talking to each other. The gap is where the machine is running.

Sources — CORP / LING Spine