Gricean Maxims Applied to Human-LLM Interaction

Summary: Kim et al. (2025) use Grice’s four cooperative maxims as a design framework for human-LLM interaction, deriving nine concrete design considerations across the three stages of the interaction cycle — and showing that each maxim needs reinterpretation to fit the specific failure modes of LLMs.

Sources: Academia/Gricean_Human-LLM.pdf

Last updated: 2026-05-06


Background

H.P. Grice’s Cooperative Principle holds that effective communication depends on interlocutors adhering to four maxims:

  • Quantity — be informative, but not more than necessary
  • Quality — be truthful; don’t say what you don’t believe
  • Relation — be relevant
  • Manner — be clear, brief, orderly; avoid obscurity and ambiguity

These maxims were formulated for human-human conversation. Kim et al. ask: how do they apply — and fail — in human-LLM interaction? The key gap is that humans assume conversational partners have inferential ability — they read contextual nuance, infer intent, and model shared knowledge. LLMs lack this in the same way; violations of the maxims emerge from structural limitations rather than communicative failure.

Reinterpretation for LLMs

Each maxim picks up an additional layer when applied to AI:

Original maximReinterpretation for LLM context
Quantity — be informativeAlso: optimize cognitive load; use hierarchy so users can drill as needed
Quality — be truthfulAlso: foster trust; hallucinations and inconsistency undermine perceived reliability, so LLMs must make accuracy and reasoning visible
Relation — be relevantAlso: track the user’s evolving goals across the conversation, not just the immediate query
Manner — be clearAlso: adapt output format (bullets, summaries, code) to the specific user and task

Two additional maxims, not in Grice, were proposed by participants: Benevolence (moral responsibility in outputs) and Transparency (explicit acknowledgment of knowledge limits and constraints). (source: Academia/Gricean_Human-LLM.pdf)

The three-stage interaction cycle

The paper maps design across three stages:

  1. User communicates a goal — user specifies task, role, preferences
  2. LLM interprets and generates — LLM plans, decomposes, produces output
  3. User assesses the output — user reviews, modifies, feeds back

Nine design considerations

DCMaxim(s)StagePrinciple
DC1Quality(2) GenerateLLMs should expose their reasoning behind outputs
DC2Quality, Relation(1) CommunicateBefore generating, LLMs should present a task decomposition plan for user approval
DC3Quality, Relation(3) AssessEnable seamless integration of user’s ongoing tasks; uninterrupted context management
DC4Quantity(2) GenerateHighlight which parts of the output the user should focus on
DC5Quantity, Relation(3) AssessEnable granular interaction — modify, adjust, provide feedback at the detail level
DC6Quantity, Manner(2) GeneratePresent output in hierarchical, expandable structure
DC7Quantity, Relation, Manner(1) CommunicateLet users preset the LLM’s role, response type, and interaction style
DC8Relation(1)+(3)Enable users to monitor and manage LLM memory — what it remembers, forgets, updates
DC9Manner(2) GenerateLLMs should prompt users for clarification when detecting ambiguity

Key insight: the Relation failure

Workshop participants identified violations of the Relation maxim as the most frustrating failure mode — LLMs that respond to the immediate query but lose track of the user’s broader goal across the conversation. DC2, DC3, and DC8 are all remedies for this single failure.

Connections

borg-llm-meaning establishes that LLMs produce meaningful outputs via derived intentionality but do not assert content. The Gricean framework adds the pragmatic layer: even meaningful outputs can violate cooperative norms. Borg asks what LLMs mean; Kim et al. ask how they should cooperate.

austin-speech-acts provides the theoretical foundation: Gricean maxims are embedded in the same pragmatics tradition. The illocutionary/perlocutionary distinction maps onto the gap the paper identifies — LLMs produce locutions but lack the cooperative commitment that makes an utterance an act.

llm-and-mind synthesizes the broader picture: LLMs are not agents and cannot infer intent the way cooperative partners are expected to. The design considerations in this paper are compensatory mechanisms for that fundamental gap.