Skip to main content

Pragmatics · 12 min

Grice's Cooperative Principle and Conversational Maxims

Speakers cooperate in conversation by following maxims of quantity, quality, relation, and manner. Apparent violations of the maxims trigger conversational implicatures: meanings inferred beyond the literal content. The foundation of formal pragmatics and the framework for understanding what an LLM does and does not communicate.

Grice's Cooperative Principle

Why This Matters

In Logic and Conversation (1975), Paul Grice proposed that linguistic communication is governed by an overarching principle: speakers cooperate in conversation. The cooperative principle factors into four maxims governing specific conversational expectations.

When a speaker appears to violate a maxim, the hearer infers an implicature: a meaning beyond the literal content. The mechanism is rational: if the speaker is being cooperative (by hypothesis) but appears to violate a maxim, there must be a reason — and that reason is the implicature.

Grice's framework is the foundation of formal pragmatics, the scaffolding for analyzing:

  • Why "Some students passed" implies "not all students passed" (scalar implicature).
  • Why "Can you pass the salt?" is a request (rather than a question about ability).
  • Why "John has three children" typically conveys "exactly three" rather than "at least three."
  • Why sarcasm and irony work (apparent quality-maxim violation that the hearer recovers).
  • Why polite indirection conveys requests gracefully.

For ML and LLM understanding:

  • LLM outputs are evaluated against Gricean expectations: a response that's overly verbose violates manner; one that refuses to answer a clear question violates relation.
  • Helpfulness, honesty, harmlessness (HHH) frameworks for RLHF map roughly onto Gricean maxims of quantity, quality, and relation.
  • LLM hallucinations are quality-maxim violations the model doesn't detect.
  • Probabilistic-pragmatics frameworks (Goodman-Stuhlmüller 2013, RSA) formalize Gricean inference as Bayesian social-cognitive reasoning.

The Cooperative Principle

Make your conversational contribution such as is required, at the stage at which it occurs, by the accepted purpose or direction of the talk exchange in which you are engaged. — Grice 1975

The principle is broad and intentionally so. The four maxims specify what it requires.

The Four Maxims

1. Maxim of Quantity

  • Make your contribution as informative as is required for the current purpose.
  • Do not make your contribution more informative than required.

Examples of quantity-related implicature:

  • "How many children do you have?" / "Three." → The hearer infers "exactly three." If the speaker had four, the maxim required saying so.
  • "Some students passed." → The hearer infers "not all students passed." If all had passed, the speaker would have said all (a scalar alternative).

2. Maxim of Quality

  • Do not say what you believe to be false.
  • Do not say what you lack adequate evidence for.

Quality-related implicature:

  • Sarcasm: "Lovely weather we're having" in a downpour. The speaker apparently violates quality; the hearer recovers the irony.
  • Hedge: "I think it's raining" implies the speaker is uncertain. Without the hedge, "It's raining" implies certainty.

3. Maxim of Relation

  • Be relevant.

Relation-related implicature:

  • "Is John dating anyone?" / "He's been visiting Paris a lot." → The hearer infers John may be dating someone in Paris.
  • Topic-shift signaling: "Anyway" and "speaking of which" flag deliberate relevance choices.

4. Maxim of Manner

  • Avoid obscurity.
  • Avoid ambiguity.
  • Be brief.
  • Be orderly.

Manner-related implicature:

  • "I drank the wine" vs "I caused some wine to enter my stomach" — the second's odd phrasing implies something unusual about the event.
  • Order matters: "They got married and had a baby" differs from "They had a baby and got married" in temporal implicature.

Implicature

A conversational implicature is a meaning the speaker conveys beyond what they literally say, recoverable through Gricean inference.

Two main types:

Generalized conversational implicature

Triggered by specific lexical items independent of context. Scalar implicature is the canonical case:

  • some → not all
  • or → not both
  • might → not certain

Levinson 2000 catalogs these and shows they are remarkably stable across languages: similar lexical scales trigger similar implicatures.

Particularized conversational implicature

Context-dependent. Same utterance triggers different implicatures in different contexts:

  • "Can I borrow your car?" / "I have to study tonight." Implicature: refusal (in this context).
  • "Did the meeting go well?" / "I have to study tonight." Implicature: change of subject (in this context).

The Cancelability Test

A defining property of conversational implicature: the implicature is cancelable. The speaker can explicitly retract it without contradiction.

  • "Some students passed — in fact, all of them did." The scalar implicature "not all" is canceled.
  • "John has three children — at least three; he might have more." The exhactness implicature is canceled.

If a meaning cannot be canceled, it is not conversational implicature but rather:

  • Entailment: logically required by the literal content.
  • Conventional implicature: triggered by specific lexical items (but implies contrast; therefore implies inference).
  • Presupposition: backgrounded content the speaker assumes.

ML Connections

LLM cooperative behavior

LLMs trained with RLHF are explicitly tuned to be:

  • Quantitatively informative (helpfulness): cover the question without padding.
  • Quality-respecting (honesty): refuse to make up facts; acknowledge uncertainty.
  • Relevant (relation): answer the question rather than changing the subject.
  • Manner-clear (clarity): avoid ambiguity, be brief, be ordered.

The HHH framework (Anthropic 2022, Helpful, Honest, Harmless) formalizes these expectations. Training-procedure refinements (constitutional AI, RLAIF, process supervision) target specific Gricean violations.

Hallucination as quality-maxim violation

LLM hallucinations are quality-maxim violations: the model states something it lacks evidence for. Detecting and mitigating hallucinations is the ML version of detecting quality-maxim violations.

Relevant techniques:

  • Retrieval-augmented generation: ground the response in a retrieved corpus, reducing the temptation to fabricate.
  • Calibration of uncertainty: train the model to say "I don't know" when it lacks confidence.
  • Process supervision (Lightman et al. 2023): score intermediate reasoning steps for factuality, not just the final answer.

Rational Speech Act framework

Goodman-Stuhlmüller 2013 formalized Gricean inference as a Bayesian process: the listener has a prior over speaker beliefs and reasons about which message a rational cooperative speaker would have sent. The framework is recursive: the pragmatic listener models the pragmatic speaker who models the literal listener.

The RSA framework reproduces classic Gricean implicatures (scalar, manner, relevance) from a Bayesian-decision- theoretic substrate. It has become a popular substrate for formal-pragmatic ML research, with extensions to dialog modeling, reference-game tasks, and prompt-engineering theory.

Probing for pragmatic understanding in LLMs

Tests of whether LLMs understand implicature exist in modern benchmarks (BIG-bench, BLiMP-pragmatic). The results: LLMs typically handle generalized scalar implicature well but struggle with context-dependent particularized implicature. The gap is one of the documented frontiers in LLM linguistic competence.

Common Mistakes

Watch Out

Conflating implicature with entailment

"Some students passed" entails "at least one student passed". The entailment is logically required and cannot be canceled. "Some students passed" implicates "not all students passed". The implicature is pragmatic and is cancelable. Entailment is truth-conditional; implicature is communicative.

Watch Out

Treating maxims as rigid rules

The maxims are defaults — expectations that hearers use to infer. Speakers regularly flout maxims (deliberately violate them to convey a message); a flouted maxim still triggers implicature. The framework is about hearer inference, not speaker constraint.

Watch Out

Assuming all cultures cooperate identically

Grice's maxims are claimed as universal but the implementations vary cross-culturally. What counts as "as informative as required" differs (Western vs East-Asian conversational norms; high-context vs low-context cultures). Cross-cultural pragmatics documents the variation; the universal claim is at the level of the principle, not the specific maxim weights.

Watch Out

Treating LLM HHH training as Gricean

The HHH framework borrows from Grice but isn't strictly Gricean. Real Gricean inference is speaker-modeling: the hearer reasons about why a rational speaker would have said what they said. RLHF tuning is reward-following: the model learns to produce outputs that score well on a reward model. The two converge in many cases but diverge in others (politeness, indirection, context-sensitivity).

Cross-Network Links

  • LinguisticsPath internal: prerequisite lambda-calculus-for-semantics for the truth-conditional substrate; relevance theory is the Sperber-Wilson alternative; presupposition and speech-act theory are the adjacent pragmatics topics to read next.
  • TheoremPath direction: RLHF, hallucination mitigation, and retrieval-augmented generation are places where Gricean framing can be useful but must not replace reward-model or retrieval mechanics.
  • EconomicsPath: the Rational Speech Act framework connects to Bayesian decision theory and signaling games in economics.
  • ComputationPath: dialog-understanding and semantics-of-natural-language pages provide formal context.

References

Canonical:

  • Grice, H. Paul. "Logic and Conversation." Syntax and Semantics, Vol. 3: Speech Acts (1975) 41-58.
  • Grice, H. Paul. Studies in the Way of Words (1989).
  • Levinson, Stephen C. Pragmatics (1983), Chapter 3.
  • Levinson, Stephen C. Presumptive Meanings: The Theory of Generalized Conversational Implicature (2000).
  • Sperber, Dan, and Deirdre Wilson. Relevance: Communication and Cognition (1986; 2nd ed. 1995).

Modern Pragmatics and Computation:

  • Horn, Laurence R. A Natural History of Negation (1989).
  • Saul, Jennifer. Lying, Misleading, and What Is Said (2012).
  • Goodman, Noah D., and Andreas Stuhlmuller. "Knowledge and Implicature: Modeling Language Understanding as Social Cognition." Topics in Cognitive Science 5 (2013) 173-184.