What do you think?
Rate this book
Language Machines: Cultural AI and the End of Remainder Humanism
How generative AI systems capture a core function of language Looking at the emergence of generative AI, Language Machines presents a new theory of meaning in language and computation, arguing that humanistic scholarship misconstrues how large language models (LLMs) function. Seeing LLMs as a convergence of computation and language, Leif Weatherby contends that AI does not simulate cognition, as widely believed, but rather creates culture. This evolution in language, he finds, is one that we are ill-prepared to evaluate, as what he terms “remainder humanism” counterproductively divides the human from the machine without drawing on established theories of representation that include both. To determine the consequences of using AI for language generation, Weatherby reads linguistic theory in conjunction with the algorithmic architecture of LLMs. He finds that generative AI captures the ways in which language is at first complex, cultural, and poetic, and only later referential, functional, and cognitive. This process is the semiotic hinge on which an emergent AI culture depends. Weatherby calls for a “general poetics” of computational cultural forms under the formal conditions of the algorithmic reproducibility of language. Locating the output of LLMs on a spectrum from poetry to ideology, Language Machines concludes that literary theory must be the backbone of a new rhetorical training for our linguistic-computational culture.
281 pages, Kindle Edition
Published June 24, 2025
Ratings & Reviews
Friends & Following
Create a free account to discover what your friends think of this book!
Community Reviews
Displaying 1 - 8 of 8 reviews
March 28, 2026
Shortly after ChatGPT3 first launched, I remember talking with a colleague who worried that a student had used AI to generate his essay. The paper exhibited many of the telltale signs of AI plagiarism at the time: sharp prose but platitudinous ideas, short paragraphs without much elaboration, no references. My colleague then copied the essay into ChatGPT3 and asked, "Did you write this essay?" The chatbot dutifully replied, "Yes, I did." I, however, was not sure if this was any meaningful kind of proof. If I ask someone whether they wrote an essay, there are a minimal set of cognitive processes I expect the respondent to execute before responding: they have to review the essay, search their memory, confirm whether they indeed wrote the essay, and then decide how to answer. "Yes" presupposes all that, even if the speaker decides to be false or evasive. ChatGPT, at that point in time, did not have such reasoning capacities, nor does it have now a global memory of all its interface usage and nor does it have access to any real-world notion of truth. Its "yes" was just part of its enthusiastic pantomime—it said yes because it was pre-trained to be compliant and upbeat. For me, this moment confirmed what Emily Bender said of LLMs: these are "stochastic parrots". They reproduce human language without any communicative intent and without any underlying ratiocination. They mimic language but they do not engage in cognition. They do not think.
Leif Weatherby does not believe that LLMs are necessarily intelligent or that they think, but he does believe that they produce authentic language. To dismiss LLMs as mere "stochastic parrots" is to miss out on what language radically is—a system of signs defined by their relations to one another. Language does not need to be grounded in real-world referents; it does not need to have intentionality. Weatherby has two major objections to how cognitive scientists frame language and language models. First, people can use words without knowing what they really mean and without connecting them to the real world (we can talk about unicorns even though unicorns do not exist; we can talk about uranium even if we do not understand anything about atomic theory or radiation, and without ever having seen uranium). Words can have meaning even if speakers do not fully understand their real-world denotation. Second, when critics try to differentiate humans from LLMs, even as LLMs demonstrate rising proficiency, they engage in a duplicitous form of goalpost-shifting: when LLMs generate language, critics insist that language must also have intent; when LLMs engage in reasoning, critics scale up the requirement of what reasoning entails. Whatever LLMs produce, critics can always a human would be more creative, more thoughtful, more reflective, more accurate. Weatherby calls this "remainder humanism" because the knee-jerk reaction to any LLM achievement is to insist that there remains some elusive human capability that has hitherto been under-appreciated—even as humans generally struggle with creativity, thoughtfulness, reflectiveness, and accuracy.
AI hallucinations are one of the most obvious failings of LLMs, but for Weatherby, they are actually not surprising at all—nor necessarily a failing. Hallucination is, in fact, the poetic function of language. Just as we can watch an actor on television say, "I now pronounce you man and wife," and still understand that a marriage has not literally occurred, so too can chatbots produce meaningful strings of words—even if they do not have a truth-value, denotation, or real-world meaning. Words do not point to things in the real world; they point to concepts and other words—they are always a simulacrum, not mediating reality itself but just other words and ideas. AI language is poetic in the Jakobsian view—a kind of computational formalism that conveys meaning. When LLMs "hallucinate," they actually reveal something about the relationship between particular words, sentences, genres—they might not be true but, based on the corpus they are modelled on, they capture something about language and genre, and the ideological underpinnings of society. While LLMs might be dismissed as predictive, it is important to note what they are predicting—when an AI classifier uses neural networks to determine if an image is a stop-sign, it is not technically classifying a stop-sign. It is predicting whether an image would be judged a stop-sign. AI reasons not about the real world but about linguistic patterns and relationships (of text to image, or text to text).
This is a simplistic and hurried distillation of Weatherby's core arguments but his book itself is highly technical and highfalutin. In a single page will appear unlikely duos: Chomsky and Kant, Frege and Derrida, Gödel and Jakobson. In part, this is because Weatherby explicitly wants to bring cognitive science and computation in conversation with semiology and humanities. In his view, the humanities and the leading humanists of the 20th century—Baudrillard, Butler, Žižek —ceded the study of "language" to scientists—Chomsky and Pinker, for example—who wanted to explain language as a cognitive phenomenon rather than a cultural one. Linguistics belonged to the sciences, semiology to the humanities, and this disciplinary division fractures the way we understand LLMs. Cognitive scientists deride LLMs as just statistical copy-machines and humanists defensively reject LLMs as computational and beyond their remit, "retreating into humanism and phenomenology." Weatherby's goal is not simply to theorize AI but to suture this deeper rupture between cognitive science and semiology—essentially, to genealogize a link from LLMs to Jakobson and to Saussure and put computation in the middle of this family chart.
Weatherby's book has a solid point: LLMs can reveal a lot about how language works and how language acquisition might occur. Chomsky famously argued that individual brains must have some universal grammar with general principles that an infant can then adapt to the specific parameters of their home language. Children, he argued, have a "poverty of stimulus"—the language they are exposed to is insufficient, fragmentary and often agrammatical (Weatherby, for some reason, repeatedly misrepresents that point, saying that Chomsky believed children need grammatical input—this is, crucially, not the case). The fact that children can infer the syntactic rules of a language without sufficient data or explicit feedback, and with often incorrect sentences, means that there must be some native instinct for the structure of language (two sentences like "the boy is hard to talk to" and "the boy is happy to help" look analogous on the surface but have a very different deep structure with different null subjects and objects—in the first the boy is the object of 'talk to' and in the second is the subject of 'help'. How could a student infer and generalize such rules without some preset understanding of the sub-surface structure of language?) Chomsky argued that such knowledge of deep structure had to have a biological origin and could not be based on statistical data. And this is beside the fact that children must also infer the meaning of these words as well. LLMs, however, by placing words on mathematical planes and mapping their proximity to one another, can be extremely effective both at modelling syntax and also at representing their meaning (for example, in their training data, an LLM will notice that 'boy' and 'girl' and 'man' often occupy similar places in the sentence and will develop probabilistic judgements about what sentences they will occupy and where).
But still, I fundamentally disagree with the thesis of this book. To return to my opening anecdote, it does matter whether LLMs reason and understand real-world referents. When I ask someone, "Did you write this essay?" I expect their answer to be more than just well-formed and superficially meaningful verbiage; I expect them to understand the question, recognize the referent of "essay", and mentally review their memory—and their answer triggers a set of social responsibilities. Depending on their answer, they may or may not be punished for plagiarism or for lying. Seeing only a binary division between ruled-based syntax and open-ended semiological systems, Weatherby entirely ignores the social and pragmatic meaning of language. Language is not simply an abstract system of signs; meaning is contractual, social and co-constructed. If I ask a student, "Did you write this essay?" and they answer instead, "well, my mother was really sick," the meaning of that utterance has little to do with the actual signs of the sentence. Both I and the student understand the repercussions of plagiarism; I know the student is losing face in this situation; I can see that the student's answer is, on the surface, not directly relevant, and I infer a much more elaborate subtextual meaning: "I didn't write the essay because I have been in difficult circumstances and I hope you can understand and give me some leeway."
The meaning of the sentence depends on shared knowledge of the social situation and requires specific cognitive steps, and it presupposes that both speakers understand each other's minds, fears and intentions, and the potential consequences of their words. In a different context with different people in different roles, a different meaning might be inferred. Imagine an interview between a journalist and an author. The interview asks, "You usually write sports essays but this one is totally different," and then adds with dramatic gushing, "Did you write this essay?" In this context, the answer "well, my mother was sick" might actually be understood not as a confession of plagiarism but as an explanation of how the writer was able to write on a totally different topic. Yes, all of this is semiotic, yes, all of this depends on an understanding of signs, and yes, an LLM might be able to reasonably predict the correct meaning in each situation. The point here, however, is that the meaning relies on two minds empathetically constructing meaning and interdependently making inferences about each other's mental state and intentions (I know that the student is giving an excuse for plagiarizing an essay; I know that an author is explaining how they came to write their essay).
Without understanding the reasoning process of LLMs, we have to be cautious about how we interpret their linguistic output—however grammatical, comprehensible, helpful, interesting, or creative it may seem. While new models might not compulsively say 'yes' to any query (as in my opening anecdote), they still might arrive at their answers in unpredictable ways and it can be difficult in different situations for readers to co-construct meaning as they normally would with human interlocutors. To give a different example, there is a huge difference between a romantic partner saying "I love you", a parent saying "I love you," and a rogue chatbot saying "I love you." In a romantic context, the words at the start of a relationship mark the beginning of a deeper intimacy (obviously people can lie or not mean what they say—but that does not change the social ritual); between parents and children, the words are the cornerstone of emotional care; for a computer, I have no idea what "I" and "love" would mean (and what such a statement would mean in the context of my browser use). The meaning of that utterance—and the commitment it entails—depends on the roles and relationships of the speakers and the mutual obligations embedded in their conversation.
Overall, I think this is provocative book, but I'm not persuaded that LLMs are more than "stochastic parrots". By pivoting to semiology, Weatherby shows how LLMs are interesting and deserve a theoretical treatment, but his book strangely ignores a lot of contemporary linguistics, and his shortchanging of pragmatics (dismissed early on as just one wrung on the ladder of reference) means that he does not fully account for how language works (and where LLMs are weird and alien). Intelligence and cognition are hermeneutic preconditions for our interactions with other humans, and while LLMs might successfully model languages, the mystery of the AI brain does shift how and whether we can interpret AI responses.
Leif Weatherby does not believe that LLMs are necessarily intelligent or that they think, but he does believe that they produce authentic language. To dismiss LLMs as mere "stochastic parrots" is to miss out on what language radically is—a system of signs defined by their relations to one another. Language does not need to be grounded in real-world referents; it does not need to have intentionality. Weatherby has two major objections to how cognitive scientists frame language and language models. First, people can use words without knowing what they really mean and without connecting them to the real world (we can talk about unicorns even though unicorns do not exist; we can talk about uranium even if we do not understand anything about atomic theory or radiation, and without ever having seen uranium). Words can have meaning even if speakers do not fully understand their real-world denotation. Second, when critics try to differentiate humans from LLMs, even as LLMs demonstrate rising proficiency, they engage in a duplicitous form of goalpost-shifting: when LLMs generate language, critics insist that language must also have intent; when LLMs engage in reasoning, critics scale up the requirement of what reasoning entails. Whatever LLMs produce, critics can always a human would be more creative, more thoughtful, more reflective, more accurate. Weatherby calls this "remainder humanism" because the knee-jerk reaction to any LLM achievement is to insist that there remains some elusive human capability that has hitherto been under-appreciated—even as humans generally struggle with creativity, thoughtfulness, reflectiveness, and accuracy.
AI hallucinations are one of the most obvious failings of LLMs, but for Weatherby, they are actually not surprising at all—nor necessarily a failing. Hallucination is, in fact, the poetic function of language. Just as we can watch an actor on television say, "I now pronounce you man and wife," and still understand that a marriage has not literally occurred, so too can chatbots produce meaningful strings of words—even if they do not have a truth-value, denotation, or real-world meaning. Words do not point to things in the real world; they point to concepts and other words—they are always a simulacrum, not mediating reality itself but just other words and ideas. AI language is poetic in the Jakobsian view—a kind of computational formalism that conveys meaning. When LLMs "hallucinate," they actually reveal something about the relationship between particular words, sentences, genres—they might not be true but, based on the corpus they are modelled on, they capture something about language and genre, and the ideological underpinnings of society. While LLMs might be dismissed as predictive, it is important to note what they are predicting—when an AI classifier uses neural networks to determine if an image is a stop-sign, it is not technically classifying a stop-sign. It is predicting whether an image would be judged a stop-sign. AI reasons not about the real world but about linguistic patterns and relationships (of text to image, or text to text).
This is a simplistic and hurried distillation of Weatherby's core arguments but his book itself is highly technical and highfalutin. In a single page will appear unlikely duos: Chomsky and Kant, Frege and Derrida, Gödel and Jakobson. In part, this is because Weatherby explicitly wants to bring cognitive science and computation in conversation with semiology and humanities. In his view, the humanities and the leading humanists of the 20th century—Baudrillard, Butler, Žižek —ceded the study of "language" to scientists—Chomsky and Pinker, for example—who wanted to explain language as a cognitive phenomenon rather than a cultural one. Linguistics belonged to the sciences, semiology to the humanities, and this disciplinary division fractures the way we understand LLMs. Cognitive scientists deride LLMs as just statistical copy-machines and humanists defensively reject LLMs as computational and beyond their remit, "retreating into humanism and phenomenology." Weatherby's goal is not simply to theorize AI but to suture this deeper rupture between cognitive science and semiology—essentially, to genealogize a link from LLMs to Jakobson and to Saussure and put computation in the middle of this family chart.
Weatherby's book has a solid point: LLMs can reveal a lot about how language works and how language acquisition might occur. Chomsky famously argued that individual brains must have some universal grammar with general principles that an infant can then adapt to the specific parameters of their home language. Children, he argued, have a "poverty of stimulus"—the language they are exposed to is insufficient, fragmentary and often agrammatical (Weatherby, for some reason, repeatedly misrepresents that point, saying that Chomsky believed children need grammatical input—this is, crucially, not the case). The fact that children can infer the syntactic rules of a language without sufficient data or explicit feedback, and with often incorrect sentences, means that there must be some native instinct for the structure of language (two sentences like "the boy is hard to talk to" and "the boy is happy to help" look analogous on the surface but have a very different deep structure with different null subjects and objects—in the first the boy is the object of 'talk to' and in the second is the subject of 'help'. How could a student infer and generalize such rules without some preset understanding of the sub-surface structure of language?) Chomsky argued that such knowledge of deep structure had to have a biological origin and could not be based on statistical data. And this is beside the fact that children must also infer the meaning of these words as well. LLMs, however, by placing words on mathematical planes and mapping their proximity to one another, can be extremely effective both at modelling syntax and also at representing their meaning (for example, in their training data, an LLM will notice that 'boy' and 'girl' and 'man' often occupy similar places in the sentence and will develop probabilistic judgements about what sentences they will occupy and where).
But still, I fundamentally disagree with the thesis of this book. To return to my opening anecdote, it does matter whether LLMs reason and understand real-world referents. When I ask someone, "Did you write this essay?" I expect their answer to be more than just well-formed and superficially meaningful verbiage; I expect them to understand the question, recognize the referent of "essay", and mentally review their memory—and their answer triggers a set of social responsibilities. Depending on their answer, they may or may not be punished for plagiarism or for lying. Seeing only a binary division between ruled-based syntax and open-ended semiological systems, Weatherby entirely ignores the social and pragmatic meaning of language. Language is not simply an abstract system of signs; meaning is contractual, social and co-constructed. If I ask a student, "Did you write this essay?" and they answer instead, "well, my mother was really sick," the meaning of that utterance has little to do with the actual signs of the sentence. Both I and the student understand the repercussions of plagiarism; I know the student is losing face in this situation; I can see that the student's answer is, on the surface, not directly relevant, and I infer a much more elaborate subtextual meaning: "I didn't write the essay because I have been in difficult circumstances and I hope you can understand and give me some leeway."
The meaning of the sentence depends on shared knowledge of the social situation and requires specific cognitive steps, and it presupposes that both speakers understand each other's minds, fears and intentions, and the potential consequences of their words. In a different context with different people in different roles, a different meaning might be inferred. Imagine an interview between a journalist and an author. The interview asks, "You usually write sports essays but this one is totally different," and then adds with dramatic gushing, "Did you write this essay?" In this context, the answer "well, my mother was sick" might actually be understood not as a confession of plagiarism but as an explanation of how the writer was able to write on a totally different topic. Yes, all of this is semiotic, yes, all of this depends on an understanding of signs, and yes, an LLM might be able to reasonably predict the correct meaning in each situation. The point here, however, is that the meaning relies on two minds empathetically constructing meaning and interdependently making inferences about each other's mental state and intentions (I know that the student is giving an excuse for plagiarizing an essay; I know that an author is explaining how they came to write their essay).
Without understanding the reasoning process of LLMs, we have to be cautious about how we interpret their linguistic output—however grammatical, comprehensible, helpful, interesting, or creative it may seem. While new models might not compulsively say 'yes' to any query (as in my opening anecdote), they still might arrive at their answers in unpredictable ways and it can be difficult in different situations for readers to co-construct meaning as they normally would with human interlocutors. To give a different example, there is a huge difference between a romantic partner saying "I love you", a parent saying "I love you," and a rogue chatbot saying "I love you." In a romantic context, the words at the start of a relationship mark the beginning of a deeper intimacy (obviously people can lie or not mean what they say—but that does not change the social ritual); between parents and children, the words are the cornerstone of emotional care; for a computer, I have no idea what "I" and "love" would mean (and what such a statement would mean in the context of my browser use). The meaning of that utterance—and the commitment it entails—depends on the roles and relationships of the speakers and the mutual obligations embedded in their conversation.
Overall, I think this is provocative book, but I'm not persuaded that LLMs are more than "stochastic parrots". By pivoting to semiology, Weatherby shows how LLMs are interesting and deserve a theoretical treatment, but his book strangely ignores a lot of contemporary linguistics, and his shortchanging of pragmatics (dismissed early on as just one wrung on the ladder of reference) means that he does not fully account for how language works (and where LLMs are weird and alien). Intelligence and cognition are hermeneutic preconditions for our interactions with other humans, and while LLMs might successfully model languages, the mystery of the AI brain does shift how and whether we can interpret AI responses.
April 26, 2026
As you delve into this review, remember to take it with a grain of salt—much of the book went over my head.
Sometimes, concepts that seem essentialy linked turn out be together for merely historical or technological reasons, and can be separated, causing a certain amount congnitive dissonance. This book argues that this is what has happened with the production of meaning and human cognition.
LLMs might not be intelligent in a human sense, but they can be used to probe the existing landscape of meaning, and generate new meaning by moving across paths in that landscape. Negating their meaning-producing capabilities, defending those capabilities as the sole province of human cognition, is an ultimately doomed rearguard action.
Viewing language primarily as a way to refer to external things and facts is too reductive. Instead, the essence of language is poetic: new meanings arise by juxtaposition and other non-referential ways of equating terms. Reference is an (admittedly useful) function built on top of that. Yeah, hallucinations are annoying allright, but they don't dent the power of LLMs as poetic machines.
Poetic machines, and ideology machines. Ideology in a wide sense, encompassing the worldview of the culture whose texts the LLM was trained on. Much has been written about how LLMs are poisoned by existing biases in society, but the flipside of that is that they are great computational tools to explore those biases.
The final part of the book is a grim prophecy. The deskilling of intellectual work that seemed unassailably "creative" is upon us. And, while we're still able for now to see the awkward seams where language machines are being grafted onto culture, those seams will likely become much more difficult to perceive in the future. So better pay attention now.
Links:
We Need To Talk About Sloppers
post
post. post
post.
Sometimes, concepts that seem essentialy linked turn out be together for merely historical or technological reasons, and can be separated, causing a certain amount congnitive dissonance. This book argues that this is what has happened with the production of meaning and human cognition.
LLMs might not be intelligent in a human sense, but they can be used to probe the existing landscape of meaning, and generate new meaning by moving across paths in that landscape. Negating their meaning-producing capabilities, defending those capabilities as the sole province of human cognition, is an ultimately doomed rearguard action.
Viewing language primarily as a way to refer to external things and facts is too reductive. Instead, the essence of language is poetic: new meanings arise by juxtaposition and other non-referential ways of equating terms. Reference is an (admittedly useful) function built on top of that. Yeah, hallucinations are annoying allright, but they don't dent the power of LLMs as poetic machines.
Poetic machines, and ideology machines. Ideology in a wide sense, encompassing the worldview of the culture whose texts the LLM was trained on. Much has been written about how LLMs are poisoned by existing biases in society, but the flipside of that is that they are great computational tools to explore those biases.
The final part of the book is a grim prophecy. The deskilling of intellectual work that seemed unassailably "creative" is upon us. And, while we're still able for now to see the awkward seams where language machines are being grafted onto culture, those seams will likely become much more difficult to perceive in the future. So better pay attention now.
Links:
We Need To Talk About Sloppers
post
post. post
post.
Read
March 12, 2026A very welcome attempt at backfilling the theorizing necessary to understand what LLM's are doing without resorting to questionable, (I'm gritting my teeth and being nice) SF-brained interpretations ("it's alive!"). A more than worthwhile read/guide for anyone hungry for serious discourse and discussion on LLM's that goes beyond keywords like "AGI" or "parrot". Enough mathematical explanation and intuition pumping to be serious, but not so much a layman will be fraught and get lost.
Very engaging and dense, some of it was a tad above my paygrade as someone who has never deeply engaged with the Frankfurt School, but who is familiar with ML, embeddings, and computing. Nevertheless, I came away feeling like I had a good understanding of Weatherby's overall points being made - both on the actual properties of language (what is it, and how does it work), what we are doing when we invoke LLMs, and some forward looking speculations based on the theorizing that was done.
Things it contributed particularly well for the discourse just off the dome: separation of language and cognition (culture-cognition connection/spectrum), the general semiotic perspective towards these systems rather than what we currently see in popular discourse, language-computation hip-joint, the high level marxian framing when useful and necessary.
Highly recommend.
Very engaging and dense, some of it was a tad above my paygrade as someone who has never deeply engaged with the Frankfurt School, but who is familiar with ML, embeddings, and computing. Nevertheless, I came away feeling like I had a good understanding of Weatherby's overall points being made - both on the actual properties of language (what is it, and how does it work), what we are doing when we invoke LLMs, and some forward looking speculations based on the theorizing that was done.
Things it contributed particularly well for the discourse just off the dome: separation of language and cognition (culture-cognition connection/spectrum), the general semiotic perspective towards these systems rather than what we currently see in popular discourse, language-computation hip-joint, the high level marxian framing when useful and necessary.
Highly recommend.
March 14, 2026
people think too much
October 29, 2025
Weatherby’s Remainder Humanism introduces an ambitious framework for understanding how AI and digital culture reshape what it means to be human. His central claim as I read him, that modern critiques of AI remain trapped within a residual humanism that privileges cognition over culture could have real force.
Yet the construction of this argument depends on a fundamental misreading of the very thinkers he marshals as examples, particularly Timnit Gebru and Emily Bender. As he puts it, “Bender has given the stochastic parrot framework a theory of meaning by subtraction[… ]The very ‘critique’ here cedes enormous power to these systems, undermining the explicit denial of meaning” (32–33). This rhetorical move reframes their ontological claim that meaning requires embodied, temporally situated interpretation into a semantic paradox of his own making, allowing him to accuse them of inconsistency while sidestepping the philosophical depth of their position.
I find it difficult to understand why Dr. Weatherby would make such a backhanded move, except that his broader argument depends on creating a vector through which he can reassert a claim already made, or more precisely, to position himself as transcending a debate he has first misrepresented.
Perhaps I’m being uncharitable, but the reframing of an ontological critique as a semantic one is the kind of category error one might expect from a first-year graduate student, not a seasoned theorist. In recasting Bender and Gebru’s ontological account of meaning as a naïve defense of human exceptionalism, Weatherby replaces a question about the grounds of meaning with one about its circulation, shifting the discussion to a semiotic register so as if all they were merely asserting is that only humans can use or produce language, thereby reinscribing the very anthropocentrism he claims to dismantle!
But that is not their critique at all.
When Weatherby writes, “The ‘stochastic parrot’ critique backhandedly confers enormous power on LLMs, theoretically depriving them of language but ceding almost mystical power to produce meaning—bad meaning, but meaning nonetheless,” he elides a crucial element of Bender and Gebru’s position: the role of the ontologically tethered human who provides the context in which meaning arises.
These so-called “stochastic parrots” do not generate meaning autonomously; meaning only emerges when a human interpreter situated in time, culture, and embodiment engages the text and reconstitutes its significance. In Peircean terms, the interpretant is not the machine but the human participant who grounds the sign within lived experience. Weatherby’s omission of this tether allows him to transform an ontological critique into a semantic paradox, creating the illusion of contradiction where none exists.
Yet the construction of this argument depends on a fundamental misreading of the very thinkers he marshals as examples, particularly Timnit Gebru and Emily Bender. As he puts it, “Bender has given the stochastic parrot framework a theory of meaning by subtraction[… ]The very ‘critique’ here cedes enormous power to these systems, undermining the explicit denial of meaning” (32–33). This rhetorical move reframes their ontological claim that meaning requires embodied, temporally situated interpretation into a semantic paradox of his own making, allowing him to accuse them of inconsistency while sidestepping the philosophical depth of their position.
I find it difficult to understand why Dr. Weatherby would make such a backhanded move, except that his broader argument depends on creating a vector through which he can reassert a claim already made, or more precisely, to position himself as transcending a debate he has first misrepresented.
Perhaps I’m being uncharitable, but the reframing of an ontological critique as a semantic one is the kind of category error one might expect from a first-year graduate student, not a seasoned theorist. In recasting Bender and Gebru’s ontological account of meaning as a naïve defense of human exceptionalism, Weatherby replaces a question about the grounds of meaning with one about its circulation, shifting the discussion to a semiotic register so as if all they were merely asserting is that only humans can use or produce language, thereby reinscribing the very anthropocentrism he claims to dismantle!
But that is not their critique at all.
When Weatherby writes, “The ‘stochastic parrot’ critique backhandedly confers enormous power on LLMs, theoretically depriving them of language but ceding almost mystical power to produce meaning—bad meaning, but meaning nonetheless,” he elides a crucial element of Bender and Gebru’s position: the role of the ontologically tethered human who provides the context in which meaning arises.
These so-called “stochastic parrots” do not generate meaning autonomously; meaning only emerges when a human interpreter situated in time, culture, and embodiment engages the text and reconstitutes its significance. In Peircean terms, the interpretant is not the machine but the human participant who grounds the sign within lived experience. Weatherby’s omission of this tether allows him to transform an ontological critique into a semantic paradox, creating the illusion of contradiction where none exists.
April 24, 2026
Dense, thoughtful work arguing for a return to Saussurean structuralism and a poetics of language. It will never cease to feel strange to me that language has passed out of our hands and is no longer fully “our own.” I have some questions about how Weatherby interprets Derrida, more on that here! https://thejester.substack.com/p/the-...
December 29, 2025
A great text on LLMs, language, culture, humanities and philosophy
some of the best thinking about LLMs I've encountered. engages with all the significant positions and stakes out interesting turf
some of the best thinking about LLMs I've encountered. engages with all the significant positions and stakes out interesting turf
February 14, 2026
This is very much a book focused on philosophy of language not mind. At times it felt tangential to even the discussion of generative AI. Nevertheless it engages well with modern philosophical understanding of language; if not really taking note of anything prior to 1800.
Displaying 1 - 8 of 8 reviews