In this article
To an AI, we speak badly. Incomplete sentences, double meanings, deixis dependent on physical context, metaphor with no marker. And yet we understand each other, between humans, almost always. Grice (1975) described how cooperative conversation rests on inference more than on literal encoding. Ambiguity isn't a bug, it's a feature: it lets us negotiate meaning on the fly, adjust nuance, say things halfway. An AI that doesn't tolerate ambiguity isn't better, it's more rigid.
The engineer and the monster
Anyone who's written code knows what the computer demands of language. Exact syntax, unique identifiers, compatible types, statement terminators, closed braces, commas in their place. A misspelled variable, an omitted semicolon, a name with a capital where a lowercase belonged, and the program doesn't run. The compiler doesn't infer intention. It reads literally what's there. What isn't there doesn't exist.
Anyone who's gone from the office to the hallway, from the hallway to the kitchen and from the kitchen to a conversation between three people who know each other, knows what human speech allows. "Pass me that thing over there." "We'll see about it tomorrow." "He's nutty as a fruitcake." "I won't say it twice." "The usual." Each of these sentences is incomprehensible to a parser. A referent is missing, context is missing, type is missing, metaphor is in excess. And the person they're addressed to understands, almost always, without blinking.
This difference between the literal pretension of code and the cooperative ellipsis of speech isn't human negligence. It's the basic mechanics of a system whose objective isn't bit-by-bit transmission. Its objective is coordination between agents who share a lot of prior structure. Speech optimizes not for literal precision but for conversational economy, and that economy demands useful ambiguity.
Grice and cooperation as infrastructure
H. P. Grice, in Logic and Conversation (1975), published the framework that opened modern pragmatics. His thesis: human conversation isn't encoding followed by decoding, it's cooperative inference. Speaker and listener assume, while the conversation works, a cooperative principle: both are trying to understand each other and to respect tacit rules —maxims— about quantity, quality, relation and manner.
The maxim of quantity asks you to provide the necessary information, no more and no less. That of quality, to say what you believe to be true. That of relation, to say what's relevant to the situation. That of manner, to say it with synthetic clarity. The interesting thing about the Gricean framework is that the apparent breach of a maxim triggers inference, not failure. If someone breaks the maxim of quantity by saying more than necessary, the listener infers some implicit meaning in the excess. If they break that of relation with a seemingly off-topic comment, the listener looks for an oblique connection. Conversation is a constant negotiation of implicatures, not a literal reading.
Sperber and Wilson, in Relevance. Communication and Cognition (Blackwell, 1986), reformulated the framework by simplifying it. They posited a single principle: in every conversation, the hearer assumes that what's said is optimally relevant —the maximum information extracted with the minimum cognitive effort—. On that assumption, the hearer infers what wasn't said. The inference is automatic, cognitively costly, and almost always right between interlocutors who share context.
Here's the detail worth holding onto. Pragmatics isn't decoration of language. It's its operating core. Without the inferential machinery of cooperation, words would remain tokens with no function. What makes them communication is the conversational calculation the interlocutors run in parallel, all the time, with no apparent effort.
Deixis and the body
There's a class of words especially revealing of the problem: deictic words. "Here," "there," "this," "that," "I," "you," "now," "tomorrow," "then." They have no fixed referent. Their meaning is computed relative to the physical, temporal and dialogic context of the moment. "Here" means a different place each time. "Tomorrow" slides every day. "You" depends on who you look at.
These words are ubiquitous in speech. Impossible to process for a system that has no body in the world. An LLM can be trained on billions of tokens where "here" appears, but "here" means nothing stable the model can learn. Each instance is a localized event. The word is a pronoun of space, and without shared space there's no referent.
The technical consequence for language models was documented by Hu and others, in A fine-grained comparison of pragmatic language understanding in humans and language models (ACL 2023). They found that LLMs solve conventional implicatures reasonably well —set phrases, standardized forms— but fail systematically at presupposition and indexical deixis, precisely where language depends on physical context not encoded in the corpus. The figure is interesting because it makes clear the problem isn't a poor corpus; it's the textual corpus per se. The world isn't learned in text alone.
Bender and Koller, in Climbing towards NLU (ACL 2020), drew attention to this structural limit. Their thesis, called the octopus hypothesis, poses a thought experiment. Imagine two people on separate islands, communicating by telegraph. An intelligent octopus intercepts the messages, learns the patterns of the exchange, and then passes itself off as one of the two. Would it fool the other? For a while, yes, on abstract topics. As soon as one of them referred to the physical world of the other island —"look at that tree," "the sun is going down," "it's cold"— the octopus couldn't sustain the conversation, because it lacks the embodied referent. LLMs, according to Bender and Koller, are the octopus: they learn the patterns of language without having access to the world the language refers to.
The metaphor with no marker
There's another layer worth mentioning because we overlook it. Everyday language is metaphorical to the marrow, and the metaphor goes unlabeled. George Lakoff and Mark Johnson, in Metaphors We Live By (University of Chicago Press, 1980), documented the phenomenon. When we talk about arguments we use the domain of war: "I defended my position," "he attacked my argument," "her thesis is indefensible." When we talk about time, we use the domain of money: "I invested hours," "I can't afford to spend time on that," "I have to save time." When we talk about affect, we use the domain of heat: "a warm relationship," "a cold reception," "human warmth."
Metaphors aren't ornament. They're conceptual structure. Lakoff and Johnson argue that thinking of time as money, in a culture, shapes the material practices around time: planning, scheduling, productivity. The metaphor steers cognition as much as it describes it.
For an LLM, the standardized metaphor is processable because it appears massively in the corpus. What's hard is the new or idiosyncratic metaphor —the one a speaker invents on the fly to capture an unprecedented nuance. That metaphor is understood between humans because both share a structure of conceptual mappings. It's understood poorly or not at all when the receiving system has only a textual corpus and not the underlying conceptual machinery.
What the LLM does when there's ambiguity
The operational question: what happens when a human user directs at an LLM a message loaded with useful ambiguity? The short answer: the model fills in with the median of the corpus. Where shared context would let a human infer one of several possible readings, the model chooses the reading statistically most common in its training distribution. If the most common reading matches the one the user had in mind, the conversation works. If it doesn't match, the model answers a different question from the one it was asked, and does so with the same fluency as if it had got it right.
This produces an operational asymmetry worth naming. The human who gets an ambiguous answer to their question notices the incongruity and translates it: "ah, it understood something else." The LLM doesn't notice the incongruity between what the user wanted and what the model understood, because it has no access to what the user wanted, only to the text sent. And the user, conversing with a fluent system, tends to read the answer as correct before verifying it was the answer to their question.
The asymmetry deepens when the user, without realizing it, adapts their question to the model. After a few misunderstandings, the user learns to write with less ellipsis, less deixis, less metaphor, more explicitness. The conversation becomes more processable but less human. The user speaks to the machine in a polished sublanguage almost nobody uses with humans. And, with prolonged use, that sublanguage starts to invade their speech with humans too. It's a recent phenomenon, anecdotally documented by sociolinguists and still without a quantitative study, but observable.
Rigidity as an industrial ideal
There's a strain of commercial discourse about AI that sells the clarity of technical language as a virtue, almost as a liberation from the "mess" of everyday speech. The line "be clear and specific when you talk to the AI" appears in any prompting guide. It's good operational advice given the system we're dealing with. It's bad advice if generalized as a description of what language should be.
Language isn't messy. It's dense. Ambiguity lets you pack meanings into few words while leaving the unpacking to cooperative inference. It's efficient, not lazy. It's flexible, not inexact. It lets you say things halfway for good reasons —because the listener holds up the conclusion better, because context constrains it, because courtesy qualifies it, because the topic requires openness—. A conversation where everything is formally specified would be unbearable and, paradoxically, less communicative.
Pinker, in The Stuff of Thought (Viking, 2007), drew attention to the sophistication of indirect speech acts. When someone says "could you pass me the salt?", syntactically they're posing a question about ability. Pragmatically they're making a request. The distance between the literal content and the effective act is enormous, and we traverse it effortlessly. Why not say it directly, "pass me the salt"? Because the indirection is a meaningful social act: it marks courtesy, distance, the symbolic optionality of the response. If it were removed, human interaction would lose a huge layer of social information.
A rigid system isn't better than a flexible one. It's more rigid. And rigidity in a domain where flexibility is the central operational property isn't a technical virtue, it's functional mismatch. A tool that demands clarity from a system built on cooperative ambiguity destroys its own effectiveness, it doesn't improve it.
The negotiation of meaning
There's a property of human speech worth naming at the end. Meaning isn't transmitted, it's negotiated. When two humans converse, they aren't sending each other clean packets of meaning between brains. They're negotiating, on the fly, a shared interpretation that each turn adjusts. The initial misunderstanding is routine, not failure. Conversational repair —"do you mean...?", "no, sorry, I meant..."— is normal machinery, not error.
This negotiation requires two things an LLM doesn't quite have. It requires a willingness to be wrong when inferring and to correct when evidence appears. And it requires a model of the other: a representation of what the interlocutor knows, what they care about, what context they're in, what they have in mind. The LLM has the first condition only partly —it agrees to rephrase when the user asks— and the second almost not at all —its model of the user is the one inferred from the prompt, with no history or biographical continuity.
The result is that human-LLM conversations are conversations impoverished in pragmatics, rich in syntax, rich in style, rich in explicit information. They work for many tasks —and here's the nuance that avoids the technophobic reading— but they don't work as a general substitute for human speech. What the LLM doesn't capture isn't vocabulary or grammar. It's the cooperative conversational apparatus that speech was built on over tens of thousands of years. Calling speech limited for not entering this apparatus is to confuse the mirror with the thing.
Definitions
Pragmatics. The branch of linguistics that studies the aspects of meaning depending on context, intention and cooperation between interlocutors, as opposed to the purely semantic or syntactic aspects.
Conversational maxim. A tacit rule of cooperative conversation, formalized by Grice (1975). The four classic ones are quantity, quality, relation and manner. Their apparent breach triggers inference, not communicative failure.
Implicature. Meaning inferred from what's said without being literally encoded in it. Human conversation is dense in conventional and particular implicatures.
Deixis. The set of words and expressions whose meaning is computed relative to the physical, temporal or dialogic context of the utterance. "Here," "you," "tomorrow" are canonical examples. They have no stable referent learnable from the textual corpus alone.
Conventional implicatures vs. presuppositions. The former are standardized inferences associated with recurring expressions. The latter are the conditions the speaker assumes true in producing an utterance. LLMs solve the former better than the latter, according to Hu and others (2023).
Octopus hypothesis. A thought experiment by Bender and Koller (2020). An intelligent octopus that learns the patterns of telegraphic messages could sustain abstract conversation but would fail as soon as the language referred to the physical world inaccessible from the corpus.
References
Bender, E. M. & Koller, A. (2020). Climbing towards NLU. On Meaning, Form, and Understanding in the Age of Data. ACL 2020. The octopus argument and a critique of equating statistical performance with semantic understanding.
Grice, H. P. (1975). Logic and Conversation. In Syntax and Semantics 3. Speech Acts, Academic Press. The foundational framework of modern pragmatics: the cooperative principle and conversational maxims.
Hu, J. et al. (2023). A fine-grained comparison of pragmatic language understanding in humans and language models. ACL 2023. An empirical evaluation of LLMs across different pragmatic phenomena; systematic failures at presupposition and indexical deixis.
Lakoff, G. & Johnson, M. (1980). Metaphors We Live By. University of Chicago Press. Documentation of metaphor as the conceptual structure underlying everyday language.
Pinker, S. (2007). The Stuff of Thought. Language as a Window into Human Nature. Viking. An analysis of indirect speech acts and the pragmatic sophistication of human language.
Sperber, D. & Wilson, D. (1986). Relevance. Communication and Cognition. Blackwell. A reformulation of Gricean pragmatics around the single principle of optimal relevance.
Comments0
No comments yet.
Leave a comment