I recently wrote a blog post in which I summarised some research on the topic of multiple meanings and AI against the backdrop of an old paper of mine on ‘ambiguities we live by’, published in 2001 in the Journal of Pragmatics.
Pragmatics is a branch of linguistics that studies how language is used in context and in social interaction and how meaning is co-constructed between interlocutors. This co-construction brings about and thrives on multiple meanings. Our varied uses of words, including metaphorical uses, can lead to words acquiring multiple dictionary meanings over time (imagine all the meanings of the verb ‘to get’ for example, ‘get it’?) – polysemy. We can exploit these multiple meanings in jokes, puns and double entendres in actual discourse; we can also misunderstand or take one word or phrase or utterance for another – ambiguity (E.g. “Did you hear about the two people who stole a calendar? They each got six months” Got it?).
In this post, I’ll focus on discourse and the co-construction of meaning in social interaction and in context. I’ll first briefly summarise my previous posts and then cite an incisive comment by my colleague Robert Dingwall which opened a whole can of pragmatic worms so to speak and deepened my reflection on AI, polysemy and ambiguity.
Previous posts
In my previous posts I engaged with various papers that discussed issues around ambiguity and polysemy in the context of AI, either how corporations exploit words with multiple meanings to manipulate users of AI or how LLMs themselves grapple with multiple meanings. As part of this, I examined a paper from 2026 that deals with the issue of ‘ambiguity collapse’.
Whereas in ordinary conversation multiple meanings can be exploited and understood productively in jokes and arguments, for example, and can even lead to social bonding (as we showed in our 2001 paper), this paper argues that LLMs cannot do this and that they destroy/collapse ambiguity or multiple meanings. They warn that this poses epistemic risks. I’d argue that this poses social risks too and I come back to that at the end of this post.
A reader comment
On 18 May 2026, Robert Dingwall, a sociologist whose work veers between ethnomethodology and symbolic interactionism, left a comment underneath my second post in which I discuss ‘ambiguity collapse’ and asks: “Have you ever looked at Larry Wieder’s chapter ‘On Meaning by Rule’ in Jack Douglas’s collection /Understanding Everyday Life/ (1971)? It’s a critical examination of ethnoscience/cognitive anthropology/Brunerian psychology for its theory of language as a set of marks that correspond to the world rather than a set of tokens that are deployed by users to create the world. The former is the behaviour of computers being fed texts. I don’t have time at the moment to develop the argument but it seems to me to be very relevant to LLMs and what they can actually achieve in practice.”
I had, indeed, not looked at that chapter. As the chapter is not readily available online, Robert send it to me and I read it. It is extremely dense (or I am no longer used to reading philosophical work!) and I’ll certainly lose some insights in my attempt to summarise it. But stay with me….
Rules, marks and tokens
Lawrence Wieder wrote his chapter “On Meaning by Rule” in 1971, a time when structural linguistics was flourishing, when ethnomethodology began to emerge, and when pragmatics, as a branch of linguistics, came on the linguistic scene. Structural linguistics looked at meaning from inside the language system, as a relationship between signs; ethnomethodology and pragmatics stepped outside the system and looked at meaning as it is produced between people in social interaction.
Wieder doesn’t mention pragmatics, but there are clear overlaps between ethnomethodological and pragmatic approaches to meaning. Pragmatics was put on the map by J. L. Austin’s 1962 book How to do Things with Words, while Harold Garfinkel launched ethnomethodology with his 1967 book Studies in Ethnomethodology. Wieder was one of Garfinkel’s first students.
In his chapter, Wieder discusses two rule-based theories of meaning linked to structural linguistics, namely the theory of ‘corresponding marks’ and the ‘theory of tokens’ and contrast them with ethnomethodological accounts of meaning. The first two theories, although slightly different, can be called ‘representational’, while the ethnomethodological account can be called ‘construcitonist’. Let’s now look at these theories.
According to the ‘theory of corresponding marks’, names correspond to sets of features with no genuine reach into a perceptual world. The word ‘uncle’ corresponds to male, one generation above, and so on (p. 133). That’s all there is. This is a binary view of meaning. The referent is nothing but the rule for correctly applying the name. In this theory it is marks and rules all the way down.
The ‘theory of tokens’ admits a real perceptual world, which is an improvement on the first. This is a triadic theory of meaning. Names are tokens that actors apply to actual objects by matching attributes against perceived features; for example, scanning for “round, red, shiny” to identify an apple (p. 117). There is a world out there and names are fitted to it by rule. Although we have now a link to perception and the world, it is still tokens and rules all the way down.
The ethnomethodological account of meaning switches our perspective away from rules. Rather than asking ‘What are the rules?’ it asks ‘How do members accomplish the activity of naming and recognising things, in this situation, at this time?’ According to the theory, criteria are not applied but elaborated on the fly – improvised, if you like; meaning is not stored and retrieved but achieved in situation/context, and the whole process is embedded in social accountability. One can get it wrong, be corrected, lose face, fall flat on one’s face, disrupt the interaction. This is not ‘meaning by rule’ but ‘meaning by interaction’. And this includes, in my view, the use and understanding of ‘ambiguities we live by’.
As Wieder points out, and every proponent of linguistic pragmatics would concur, “in everyday talk persons constantly use expressions the sense of which is relative to the place in which it is spoken, what the hearer knows about the speaker, the time at which it is spoken, and an indefinitely extendable collection of other contextual matters.” (p. 108) Meaning construction between interlocutors is situated, flexible and adaptable. And going beyond Wieder and back to my posts: In ordinary language ambiguity is a feature, not a bug.
What has all this to do with AI? As we have seen in the posts, most of the articles I discussed see ambiguity in a negative light apart from one outlier that looks more closely at the inner workings of LLMs and suggests that ambiguity might be accommodated in LLMs. In Wieder’s framework accommodating a flexible use of ambiguity or multiple meanings into understanding LLMs would be difficult, as we shall see.
Puppetry and computers
The crunch of the 1971 paper comes when Wieder uses insights from the sociologist Alfred Schütz, introduces the fascinating concept of ‘puppetry’, and maps the marks-based approach to meaning onto computers (p. 121-122).
According to Schütz, when working on a sociological theory, one can construct a theoretical ‘puppet’ that does nothing other than what a theory proposes, and then ask: what kind of society would these puppets inhabit? What would their world look like? The method is designed to expose what a theory quietly assumes or rules out.
When Wieder applies this to the marks/tokens view of meaning the puppet society that emerges is rather strange. Members are essentially “courses of talk” (p. 122) and nothing more. Object features consist only in what members verbally attribute. No lies are possible, because with no perceptual world as reference; whatever a member says is just so: “I said it, therefore it is.” (see p. 123) There are no perspectives, no here and there (no indexicality), no genuine intersubjectivity, just simultaneous streams of talk.
Wieder’s point is that this puppet society is exactly the society a computer inhabits. He says so explicitly! The theory of meaning based on corresponding marks is “fully compatible with the built-in presuppositions of the computer, which treats all inputs as marks.” (p. 124)
He wrote this in 1971, but it reads now like a description of an LLM. (The reference to a computer comes completely out of the blue, by the way, as computers are not mentioned anywhere else in the chapter!) He also points out that: “This is the same as treating the world you are attempting to represent from the stance of a blind man who has been kept in a straightjacket and who has no way of addressing his world except by means of his hearing and speaking.” (p. 124) (This
So, Wieder argued in 1971 that a theory of corresponding marks produces puppet societies that no actual member would recognise and that this puppet world is precisely the world a computer inhabits. In terms of modern AI systems that means: LLMs have stored immense amounts of marks and learned to carry out sophisticated pattern matching, but they don’t deploy these marks in open-ended social situations with other people around them. LLMs are, in Wieder’s terms, the most sophisticated marks-processors ever built, but that’s all.
One might think that the ‘theory of tokens’ lets LLMs off the hook. After all, they are trained on vast amounts of language used in real contexts and have absorbed something like a statistical shadow of how words are fitted to situations. But Wieder closes this route. Even in the token theory, the actual process by which actors fit names to objects, i.e. the situated, accountable, correctable work of meaning-making, is treated as theoretically invisible. What matters is only the product, not the method (see p. 126). And that covertness is what disqualifies LLMs too. However sophisticated their outputs, the fitting process is opaque, unaccountable, and not open to correction by a co-present and co-producing other in the way that human meaning-making always is.
Why Wieder matters
After all this we can come back to the last part of my second post and a chapter by Gur-Arieh on ‘ambiguity collapse’ and productive ambiguity. In this context Wieder’s chapter matters the most.
Productive ambiguity, as when my husband quips “I thought they had folded” upon seeing a sign for a meeting of the Origami Society, only exists between people in social interactions, people that co-construct meaning against the background of a situation and using a whole lot of pragmatic insights. That is what makes polysemy ‘live’ rather than ‘dead’ and, with it, language itself. As LLMs have no ‘situation’ or ‘other people’ to play with, they seem to collapse ambiguity prematurely (as Benjamin Riley recently said, they are “dead metaphor devices”). Collapsing ambiguity in social situations would matter to people but it doesn’t matter to LLMs.
This LLM blindness to situated action, polysemy and ambiguity is not just a trivial state of affairs; it can have consequences for science and society. This is in part explored in Victor Capraro’s recent preprint on “LLMorphism” or “the biased belief that human cognition works like a large language model”. Capraro argues that we might increasingly think about ourselves, and especially our ‘thinking’, as large language models and that such a dominant metaphor might have serious consequences.
Capraro claims that “LLMorphism blurs the distinction between saying and knowing” and that this could “contribute to a broader epistemic shift: from evaluating whether claims are grounded, justified, and accountable, to evaluating whether they are coherent, fluent, and plausible.”
In Wieder’s terms, we might begin to think of ourselves as mere ‘courses of talk’ exchanging marks or tokens, rather than situated and embodied human beings co-constructing meaning in social interaction. That’s dangerous!
Image: Bodleian: Illuminated border depicting a puppet show, 1338–1344; Flemish illuminator Jehan de Grise and his workshop
Making Science Public 
Leave a comment