Semantics is the study of what words and sentences mean by convention; pragmatics is the study of what speakers mean by using those words in a particular situation. A sentence like “It’s cold in here” has a fixed semantic content about temperature, but a listener who then closes the window has understood something pragmatic: the speaker was making a request. This division sounds clean in theory, but the boundary between the two has generated decades of heated debate among linguists and philosophers, and the way human brains actually process meaning suggests the two systems are deeply entangled.
What Semantics Handles
Semantics deals with the meanings that expressions carry regardless of who says them, when, or why. The word “cat” refers to a certain kind of animal whether you whisper it at midnight or shout it in a crowded room. A sentence like “Every student passed the exam” has a truth condition you can evaluate without knowing anything about the speaker’s intentions or mood. Semantics covers the rules by which word meanings combine into sentence meanings: the way “big red ball” means something different from “red big ball” (one sounds natural, the other doesn’t) is a semantic fact about English.
Crucially, semantics also handles things like logical relationships between sentences. If someone says “All mammals are warm-blooded” and “A whale is a mammal,” semantics lets you derive “A whale is warm-blooded” without any information about the conversational setting. These are meaning relationships that hold by virtue of the language system itself.
What Pragmatics Handles
Pragmatics picks up where the dictionary and grammar leave off. It covers the inferences listeners draw based on context, shared knowledge, and assumptions about what the speaker is trying to accomplish. If you ask a friend “Did you enjoy the movie?” and they reply “Well, the popcorn was good,” the semantic content of their answer is about popcorn. The pragmatic meaning is that they didn’t enjoy the movie much. Nothing in the sentence’s literal meaning says anything negative about the film, but you arrive at that conclusion effortlessly.
The philosopher Paul Grice laid much of the groundwork here by proposing that conversations are governed by cooperative expectations. Speakers are expected to be informative, truthful, relevant, and clear. When someone appears to violate one of these expectations, listeners don’t usually assume the speaker has gone off the rails. Instead, they search for an implied meaning that would restore the expectation. That search is the engine of pragmatic inference.
The Boundary Problem
The division between semantics and pragmatics is one of the most contested questions in the philosophy of language. On one side are semantic minimalists, who argue that a sentence’s semantic content is a complete, truth-evaluable proposition formed by combining the syntactic structure with the conventional meanings of its words. On the other side are contextualists, who maintain that context often needs to step in before a sentence expresses anything truth-evaluable at all.
Consider “She is tall.” The semantic minimalist says this sentence expresses a minimal proposition about some female individual being tall, and that proposition is either true or false even before you know whether “she” refers to a basketball player or a five-year-old. The contextualist objects that “tall” is meaningless without a comparison class supplied by context, so the sentence doesn’t express a complete thought until pragmatics fills in the gap. Recent philosophical work has defended the minimalist position, arguing that minimal propositions are truth-evaluable and that the apparent incompleteness contextualists point to doesn’t actually prevent the sentence from having a determinate truth value.
1Journal of Education, Humanities and Social Sciences. Semantic Minimalism’s Truth-Evaluability: The Minimal Proposition and Semantic UncertaintyThis debate isn’t merely academic hairsplitting. Where you draw the line between semantics and pragmatics affects how you build language technologies, how you diagnose language disorders, and how you think about legal interpretation. If the “plain meaning” of a statute is a semantic question, then a minimalist and a contextualist will disagree about what counts as plain meaning.
Scalar Implicature and the Cost of Going Beyond Literal Meaning
One of the best-studied examples of the semantics-pragmatics interface is scalar implicature. If someone says “Some of the students passed,” the semantic meaning is compatible with all of them passing. But listeners almost always infer that not all of them passed, because if the speaker had known that all passed, they would have said “all.” The word “some” sits on a scale with “all,” and listeners use pragmatic reasoning to infer that the stronger term doesn’t apply.
This process turns out to be mentally taxing. A systematic review of research on the topic found that deriving scalar implicatures draws on working memory, supporting what’s called the “literal-first” account: your brain initially processes the semantic (literal) meaning and then expends additional cognitive effort to arrive at the pragmatic interpretation.
2PubMed. Some scales require cognitive effort: A systematic review on the role of working memory in scalar implicature derivation This finding has been replicated across different testing methods, languages, and age groups, and the processing cost is robust enough that researchers now treat it as an established phenomenon rather than an open question.
3Frontiers in Communication. Sources of cognitive cost in scalar implicature processing: A reviewWhat remains debated is the source of that cost. Is it the inference itself that’s expensive, or is it related to other logical operations (like suppressing the literal meaning) that happen to coincide with the inference? The answer matters because it shapes how we understand the architecture of language processing. If the inference itself costs effort, it suggests pragmatic reasoning is a genuinely separate cognitive step layered on top of semantic decoding. If the cost comes from something else, the boundary between the two might be blurrier in the brain than it is in theory.
4Frontiers in Communication. Sources of cognitive cost in scalar implicature processing: A reviewWhat Brain Imaging Reveals
When researchers scan people’s brains while they process pragmatic meaning, a distinctive pattern emerges. In one fMRI study, participants listened to conversational exchanges containing indirect replies, where someone’s intended meaning diverged from what they literally said. Compared with direct replies, indirect replies activated a broad network spanning both the language system (regions in the frontal and temporal lobes on both sides of the brain) and the theory-of-mind network, which is the set of regions involved in reasoning about other people’s mental states.
5PubMed. Effects of contextual relevance on pragmatic inference during conversation: An fMRI studyThis is revealing because it shows that understanding pragmatic meaning isn’t just harder language processing; it recruits a qualitatively different cognitive system. You’re not only decoding words, you’re modeling what the other person knows, believes, and intends. The study also found that when replies were irrelevant (not just indirect but seemingly off-topic), a gradient of activation appeared in the right temporal lobe, as though the brain was working progressively harder to find a connection as relevance decreased.
6PubMed. Effects of contextual relevance on pragmatic inference during conversation: An fMRI studySeparate imaging work on metaphor and sarcasm shows that non-literal language engages still other specialized regions. Metaphor processing activated a structure deep in the brain associated with selecting among candidate meanings, while sarcasm specifically engaged the amygdala, a region central to social and emotional processing. Both metaphor and sarcasm activated areas linked to mentalizing, the process of thinking about what’s going on in someone else’s mind.
7PubMed. Distinction between the literal and intended meanings of sentences: a functional magnetic resonance imaging study of metaphor and sarcasmThe picture that emerges is that semantic processing and pragmatic processing share real estate in the brain’s language network but that pragmatics also recruits social cognition regions that pure semantic processing does not. The two systems talk to each other constantly, but they’re not the same system.
How Children Learn to Go Beyond Literal Meaning
Children become semantically competent before they become pragmatically competent, and the gap between the two tells us something important about how the distinction works in practice. Infants first develop phonological skills (recognizing and producing sounds), then build a vocabulary, then learn to combine words into sentences. All of this is essentially the semantic pipeline: connecting sounds to meanings and meanings to structures. Pragmatic competence, the ability to use language appropriately in social context, comes later and develops more gradually as children refine both their linguistic and non-linguistic communication skills.
8PubMed Central. Pragmatic Language Development: Analysis of Mapping Knowledge Domains on How Infants and Children Become Pragmatically CompetentThis developmental order makes sense when you consider the brain imaging findings. Pragmatic reasoning depends on theory-of-mind abilities, and theory of mind itself undergoes significant development throughout early childhood. A three-year-old who can use the word “some” correctly in a sentence may still interpret “some” literally and not draw the scalar implicature that “not all” is implied. Children gradually learn to read conversational context, detect irony, and understand that what someone says and what they mean can be two very different things.
The scalar implicature literature supports this directly. Since deriving these inferences taxes working memory, young children with limited working memory capacity often default to literal interpretations. As their cognitive resources develop, they start to behave more like adults, enriching literal meanings with pragmatic inferences. This is one reason children sometimes seem oddly literal: they’re not misunderstanding the words, they’re just not yet performing the extra cognitive work that adults do automatically.
When Pragmatic Ability Breaks Down
The distinction between semantics and pragmatics becomes clinically visible in certain developmental conditions. Children on the autism spectrum often have intact or near-intact semantic knowledge, meaning they understand words and can construct grammatical sentences, but they struggle with pragmatic aspects of language: taking turns in conversation, understanding implied meaning, interpreting non-literal speech, and adjusting their language to the social context.
A study comparing children evaluated for autism spectrum disorder with children who did not receive that diagnosis found that the ASD group scored lower on a standardized measure of general pragmatic ability, even after accounting for differences in structural language skills.
9PubMed Central. Structural and Pragmatic Language Impairments in Children Evaluated for Autism Spectrum Disorder (ASD) This dissociation, where structural language (the semantic and syntactic machinery) is relatively preserved but pragmatic language is impaired, is one of the hallmarks of ASD communication profiles.
Pragmatic language impairment can also appear in children without ASD. Social (pragmatic) communication disorder is a diagnosis specifically for cases where pragmatic skills are impaired in the absence of the restricted interests and repetitive behaviors characteristic of autism. The existence of this diagnosis underscores that pragmatic ability is a distinct dimension of language competence, not merely a byproduct of general intelligence or vocabulary size. You can know what every word in a joke means and still completely miss the joke.
Can Machines Do Pragmatics?
Large language models like GPT-4 have become remarkably good at semantic tasks: they can parse sentences, answer factual questions, and generate grammatically correct text. But pragmatics has proven to be a much harder nut to crack. A benchmark dataset called PUB (Pragmatics Understanding Benchmark) tested nine language models across four categories of pragmatic phenomena: implicature, presupposition, reference, and deixis. The results showed a noticeable gap between human performance and model performance, with huge variation across different pragmatic tasks.
10Findings of the Association for Computational Linguistics. PUB: A Pragmatics Understanding Benchmark for Assessing LLMs’ Pragmatics CapabilitiesA separate benchmark, the Contextual Emotional Inference (CEI) test, specifically targets the ability to disambiguate pragmatically complex utterances, situations where the intended meaning depends heavily on context and emotional cues. The researchers describe pragmatic reasoning as “inferring intended meaning beyond literal semantics” and note that it “remains difficult for large language models.”
11arXiv. CEI: A Benchmark for Evaluating Pragmatic Reasoning in Language ModelsThere are bright spots. When tested on interpreting the gradable adjective “strong” in different contexts, some LLMs could produce human-like distributions of meaning that shifted appropriately with context, a genuinely pragmatic feat. But they struggled with negation, a seemingly simple operation that humans handle without much trouble in pragmatic contexts.
12arXiv. Evaluating statistical language models as pragmatic reasonersOn the theoretical side, the Rational Speech Act (RSA) framework has emerged as the most influential computational model of pragmatic reasoning. RSA treats communication as a kind of probabilistic game in which a speaker chooses utterances to be informative and a listener reasons backward from the speaker’s choice to infer what they meant. The framework integrates linguistic structure, world knowledge, and context, and it has made successful quantitative predictions about human behavior across a range of tasks.
13Annual Review of Linguistics. The Rational Speech Act Framework RSA has been used to explain phenomena like hyperbole and vagueness, cases where the literal semantic content of what someone says dramatically undersells or oversells the intended message.
14Trends in Cognitive Sciences. Pragmatics through probabilistic interpretationThe gap between LLM performance and human pragmatic ability is a useful diagnostic for the semantics-pragmatics distinction itself. If the two were really the same thing, a system that had mastered semantic understanding would automatically be good at pragmatics. The fact that LLMs can handle one far better than the other suggests they are genuinely different cognitive tasks.
Gestures, Tone, and Meaning Beyond Words
Pragmatic meaning doesn’t travel through words alone. Hand gestures, facial expressions, and vocal emphasis all carry pragmatic information, sometimes overriding the semantic content of what’s said. Beat gestures, the small rhythmic hand movements speakers make to emphasize certain words, are a striking example. They don’t have any semantic content of their own. You can’t look up the meaning of a hand flick the way you look up a word. But they profoundly affect how listeners interpret what they hear.
A recent study found that when beat gestures were combined with prosodic emphasis (stress on a word), listeners were more likely to accept factually incorrect statements as true, a phenomenon called semantic illusion. The researchers interpreted this as evidence that beat gestures aren’t just spotlights for important words. Instead, they automatically trigger inferences about the speaker’s confidence, leading listeners to commit more fully to the truth of what’s being said.
15PubMed. Beat gestures and prosodic prominence interactively influence language comprehension This connects directly to Grice’s cooperative principle: listeners assume the speaker is being truthful and use every available cue, linguistic and bodily, to evaluate that assumption.
This multimodal dimension of pragmatics is something that text-based AI systems miss entirely. When you read a transcript, you lose the pragmatic information carried by gesture and tone. A sentence typed in a chat window can be ambiguous in ways it never would be face to face, which is why people invented emoji and “/s” tags. Those are crude substitutes for the rich pragmatic signaling that bodies and voices do naturally.
Pragmatics Before Language
One of the more provocative ideas in recent research is that pragmatics might be evolutionarily older than semantics. In animal communication, signals don’t carry fixed dictionary-style meanings the way human words do. Instead, the meaning of a signal for the animal receiving it depends on integrating the signal with the social context: who sent it, what was happening at the time, and what the receiver already knows about the sender. In long-lived social species where the same individuals interact repeatedly, communication becomes a system of pragmatic inference in which meaning depends on perception, memory, and social knowledge.
16Elsevier / Animal Behaviour. The origin of meaning in animal signalsThis perspective suggests that in the human lineage, pragmatic reasoning served as a precursor to the later evolution of semantics and syntax.
17Elsevier / Animal Behaviour. The origin of meaning in animal signals In other words, our ancestors were reading context and inferring intentions long before they had a system of words with stable, conventional meanings. Language evolved to plug into an already-existing pragmatic infrastructure, not the other way around. If this is right, it reframes the semantics-pragmatics distinction in a fundamental way. Semantics isn’t the foundation on which pragmatics was built; pragmatics is the foundation on which semantics was built.
This idea resonates with what we see in child development, where babies engage in pragmatic-like communicative behavior (pointing, gaze-following, reaching) before they have any words at all. It also fits the clinical picture in autism, where the semantic system can be intact while the evolutionarily older pragmatic system is impaired, as though the newer and older systems can come apart. Whether you find this evolutionary narrative fully convincing or not, it’s a useful reminder that the tidy textbook hierarchy, where semantics comes first and pragmatics is added on top, may have things backward.
Deixis and the Leaky Border
Some categories of words live squarely on the border between semantics and pragmatics, and deictic expressions are the most notorious example. Words like “this,” “that,” “here,” “there,” “now,” and “yesterday” have stable dictionary entries, so they have semantic content. But their reference shifts completely with the situation of utterance. “Here” means a different place every time you say it. “Yesterday” means a different day. These words are semantic shells that pragmatics fills with content.
Research on deixis in Yucatec Maya has argued that the basis of deictic reference isn’t even spatial proximity, as many linguists have assumed, but rather the perceptual, cognitive, and social access that participants have to the referent.
18Elsevier / ScienceDirect (Journal of Pragmatics). Fieldwork on deixis In other words, “this” versus “that” might not be about physical closeness so much as about whether the speaker and listener can readily perceive or attend to the thing being pointed out. That makes deixis an inherently pragmatic phenomenon dressed up in semantic clothing.
Deixis also turned out to be one of the pragmatic categories where LLMs showed the most variable performance, which makes sense. A machine reading a transcript has no body, no spatial location, and no perceptual access to anything. The deictic words are there in the text, but the pragmatic infrastructure needed to resolve them is missing. Handling deixis well requires exactly the kind of situated, embodied context that separates human communication from text processing.