💬 NLP

Pragmatics beyond humans: meaning, communication, and LLMs

This paper argues that the emergence of large language models necessitates a reconceptualization of pragmatics beyond traditional human-centric frameworks, proposing instead a Human-Machine Communication approach that addresses anthropomorphic biases, adopts probabilistic models like Rational Speech Act, and accounts for the paradox of "context frustration" in generative AI interactions.

Original authors: Vít Gvoždiak

Published 2026-08-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Vít Gvoždiak

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Language is more than just a string of words arranged by grammar rules. It is a social tool, a way for people to act together, share intentions, and navigate the complex, unspoken rules of a conversation. For decades, linguists have tried to map how meaning works by breaking it down into three layers: the structure of the sentence, the literal definition of the words, and the context in which they are spoken. This third layer, known as pragmatics, is where the real magic of human communication happens. It is the difference between a statement of fact and a request, or between a joke and an insult, depending entirely on who is speaking, when, and why. Now, a new kind of machine has entered the conversation. Large language models, the powerful artificial intelligence systems that can write essays, answer questions, and hold conversations, are forcing scientists to rethink these old maps. The central question is no longer just whether these machines can mimic human speech, but whether they understand the social dance of meaning at all, or if they are simply predicting the next word in a sequence without ever grasping the intent behind it.

A researcher named Vít Gvoždiak has taken a close look at how we study these machines and the theories we use to judge them. He argues that our current methods are flawed because they try to force a square peg into a round hole. For a long time, the standard way to evaluate these models has been to treat them like humans. Researchers ask them to perform tasks that require understanding social cues, like recognizing an apology or detecting sarcasm, and then compare their answers to what a person would say. If the machine gets it right, it is often praised for having "pragmatic competence." If it gets it wrong, it is said to lack understanding. Gvoždiak suggests this approach is misleading. It assumes that because the machine produces human-like text, it must be operating on human-like principles. But these machines are not human; they are statistical engines trained on vast amounts of internet text. They do not care about truth or social cooperation in the way people do; they care about probability. They are designed to guess which word is most likely to come next, not to communicate a shared reality.

The paper proposes that we stop trying to fit these machines into human categories and instead look at the unique way they interact with us. Gvoždiak suggests a new framework called Human-Machine Communication. This approach recognizes that when a person talks to a chatbot, they are not just talking to a tool, nor are they talking to a person. They are in a hybrid space where both sides shape the conversation in ways that have never existed before. The machine is not a passive recipient of commands, and the human is not just a user. They are co-conspirators in a new kind of dialogue. The author points out that much of the current research suffers from a problem he calls "substitutionalism." This happens when scientists take the results from one specific model, or one specific language like English, and assume it applies to all artificial intelligence and all human languages. Even more subtly, researchers often swap out the human participant in a conversation with a machine, assuming the machine can stand in for a person. This blinds them to how humans actually change their own behavior when they talk to a machine. People are starting to speak differently, using phrases and structures that mimic the AI, creating a feedback loop where the machine influences the human just as much as the human influences the machine.

One of the most striking findings in the paper is the concept of "context frustration." In human conversation, context is the shared background knowledge that allows us to understand each other. If I say "It's cold in here," you know I might be asking you to close the window. With a machine, this shared background is missing. The machine has no personal history, no physical body, and no real-world experiences. To make the machine understand, humans are forced to do extra work. They have to write long, detailed prompts, specifying exactly who the machine should pretend to be, what the setting is, and what the goal of the conversation is. This is a phenomenon known as prompt engineering. The irony is that while the technology is getting better at processing more and more information—remembering longer conversations and reading larger documents—the human user feels more frustrated. The machine's ability to handle vast amounts of data has not solved the problem of shared understanding; it has made the gap between human and machine context feel wider. The user is constantly trying to build a bridge of context for the machine, only to find that the machine's understanding is still fundamentally different from their own.

The paper concludes that we need to stop asking if machines can be human and start asking how they are different. The old theories of language, which were built for human-to-human interaction, are not enough to explain what is happening now. We need new ways of thinking that acknowledge the machine's unique nature. Instead of judging a machine by whether it can pass a test designed for a person, we should study the unique, hybrid conversations that happen between people and machines. These interactions are creating new forms of meaning, new social rules, and new challenges. The author suggests that the future of language study lies in understanding this partnership, where humans and machines are constantly adjusting to each other, creating a shared space that is neither fully human nor fully mechanical, but something entirely new. The goal is not to make the machine more like us, but to understand how we are changing in its presence.

Technical Summary: Pragmatics beyond humans: meaning, communication, and LLMs

Problem Statement
The paper addresses the theoretical and methodological inadequacies of traditional pragmatics when applied to Large Language Models (LLMs). Conventional pragmatics is often conceptualized as a subordinate, third dimension of meaning within a hierarchical semiotic trichotomy (syntax $\rightarrow$ semantics $\rightarrow$ pragmatics), grounded in human-specific assumptions such as Gricean cooperative principles, truth-conditional evaluation, and biological intentionality. The author argues that this framework fails to account for the connectionist, probabilistic nature of LLMs. Specifically, current research suffers from:

Anthropomorphic Bias: Applying human-centric theories (e.g., speech act theory based on truth and intention) to systems that operate on word-occurrence probabilities rather than truth conditions.
Substitutionalism: The uncritical substitution of specific models for "LLMs" in general, specific languages for "language" in general, and human subjects for LLMs in experimental designs, thereby obscuring the hybrid nature of human-machine interaction.
Contextual Paradox: The tension between the massive expansion of technical context (larger context windows, training data) and the simultaneous "collapse" of shared human context, leading to a state the author terms "context frustration."

Methodology
The paper employs a conceptual and methodological analysis rather than empirical experimentation. It proceeds through four interconnected sections:

Theoretical Re-framing: It critiques the traditional semiotic trichotomy (Morris/Carnap/Kaplan) and proposes the Human-Machine Communication (HMC) framework (Guzman & Lewis) as a superior alternative. This framework analyzes communication through functional, relational, and metaphysical dimensions rather than discrete semiotic levels.
Theoretical Alignment: It contrasts human-centred pragmatic theories (Gricean) with Probabilistic Pragmatics, specifically the Rational Speech Act (RSA) framework. The author argues that RSA, which models communication as a recursive, utility-optimizing process based on probability distributions, aligns better with the teleology of LLMs (optimization of next-word prediction) than truth-evaluation models.
Critical Analysis of Research Practices: It identifies and categorizes three forms of substitutionalism in current literature:
- Generalizing: Treating specific model outputs as representative of all AI.
- Linguistic: Extrapolating English-centric findings to all languages.
- Communicative: Replacing human interlocutors with LLMs in experimental designs, ignoring how human behavior adapts to AI.
Conceptual Synthesis: It introduces the concept of "context frustration" to describe the paradox where increased technical context (tokens, data) fails to resolve the fundamental incommensurability between human and machine presuppositions, leading to a collapse of shared context and a demand for users to co-construct pragmatic conditions.

Key Contributions

Reconceptualization of Pragmatics: The paper argues that pragmatics should not be viewed as a hierarchical "wastebasket" for meaning but as a dynamic interface where language operates as a socially embedded tool for action. It posits that the boundaries between syntax, semantics, and pragmatics are blurred in connectionist architectures.
Adoption of the HMC Framework: It advocates for shifting from a human-centric view to an HMC perspective that acknowledges machines as active communicative entities, analyzing the functional, relational, and metaphysical implications of this shift.
Proposal of Probabilistic Pragmatics: It suggests that the Rational Speech Act (RSA) framework offers a more methodologically sound approach for LLMs because it focuses on optimization and probability rather than truth conditions and human rationality.
Identification of Substitutionalism: The paper formally defines "substitutionalism" as a methodological flaw in LLM research that creates asymmetries by substituting specific models, languages, or human roles, thereby distorting the evaluation of both machine capabilities and human behavioral changes.
Definition of "Context Frustration": It introduces a new term to describe the experiential and structural dissonance in human-LLM interactions, characterized by the failure of massive technical contextualization to achieve genuine contextual alignment, forcing users into a cycle of co-constructing pragmatic conditions.

Results and Findings
As a theoretical paper, it does not present experimental data but derives several critical findings:

Destabilization of Semiotic Hierarchy: The connectionist nature of LLMs destabilizes the traditional view that syntax is the base, semantics the middle, and pragmatics the top. Pragmatic effects in LLMs manifest across all levels of meaning-making as continuous, graded processes rather than discrete, post-propositional inferences.
Inadequacy of Gricean Models: Traditional Gricean approaches, which rely on the assumption of cooperative rationality and truth-seeking, introduce methodological bias when applied to LLMs, which are designed to maximize distributional probabilities.
Human Behavioral Adaptation: Current research often ignores how human communicative behavior changes in hybrid interactions. Evidence suggests humans are beginning to mimic LLM linguistic patterns (e.g., in spoken language) and engage in "prompt engineering," adapting their own communicative strategies to the model's affordances rather than vice versa.
Unattainable Equilibrium: Unlike human-to-human communication where contexts tend to converge toward equilibrium, human-LLM interactions remain contextually unstable and defective because the sets of presuppositions and possible worlds available to LLMs are incommensurable with those of humans.

Significance
The paper claims that pragmatic theory requires significant adjustment or expansion to account for communication involving generative AI. Its significance lies in:

Methodological Correction: It provides a framework to move beyond the "anthropomorphic bias" that currently dominates LLM evaluation, suggesting that researchers must stop treating LLMs as failed humans and start analyzing them as distinct communicative agents.
Theoretical Evolution: It challenges the "wastebasket" definition of pragmatics, proposing instead that pragmatics is a dynamic interface essential for understanding the operation of language as a tool for action in a mixed human-machine ecosystem.
Addressing the Context Paradox: By introducing "context frustration," the paper highlights a critical structural limitation in current AI-human interaction that cannot be solved merely by increasing model size or context window, pointing to the need for new theoretical tools to understand the co-construction of meaning in hybrid environments.

The paper concludes that the entry of LLMs into communicative practices necessitates a shift from viewing communication as an exclusively human process to recognizing it as a hybrid performance involving machines, requiring a re-evaluation of the roles, contexts, and theories that define pragmatic inquiry.

More like this