Critical Inquiry Critical Inquiry

Leif Weatherby. Language Machines: Cultural AI and the End of Remainder Humanism. Minneapolis: University of Minnesota Press, 2025. 264pp. 

Review by Matthew Kirschenbaum

24 December 2025

There is a moment halfway through Leif Weatherby‘s Language Machines—during a discussion of Ferdinand de Saussure on the sign, which is itself bracketed by a discussion of semiology and what Weatherby terms the “semiological surround” of contemporary Western life, which is in turn grounded in the previous chapters’ spadework covering broad swaths of twentieth-century linguistic theory—when he offers, almost as an aside: “Language is not thought or meaning as such, but I cannot convey this fact outside of language” (p. 103).

Fuckin’-A. This rich, dense, frankly difficult, and conceptually challenging book about generative artificial intelligence is a self-contained testament to that truth, a truth that is true of all books of course but a book about language and instruments of language can’t help but be meta in its remit to filter the author’s thinking and thought through the scrim of prose—which a handful of diagrams and illustrations aside, is all the reader has of Weatherby on the page or screen in front of them. That he is surely in on the joke does not make the going any easier.

Language Machines is the first extended treatment of generative AI to address one of its most prominent constituent types—Weatherby would say its elemental type—on its own terms, and that is language. Not writing as input or output, but language. And language not as some ineffable reserve that is the castle keep of the remainder humanism of the subtitle but language as a field of inquiry, arguably (as many did argue) the governing inquiry for much of the twentieth century. Above all Weatherby is concerned to knock language off of its referential pedestal, which is to say that so-called communicative intent is finally only a second order property of the sounds we shape with our mouths and the symbols we scratch or tap—which are now also generated through autonomous mathematical processes. (When I f-bombed the previous paragraph, were you really just concerned to grasp my communicative intent?)

The argument for these positions is developed over the book's first five chapters in a technical forensics about the commonalities of language and computation. I say technical here less in the sense of technology as such (though we do get into actual technologies) and more in the way a rock climber might use the term: the route is technical, it demands precision, focus, and grit. Chapter 1 is an orientation to the predominant schools of linguistic thought in the twentieth century, including the syntactic (Noam Chomsky of course), statistical/empirical (exemplified by J. R. Firth), and structuralist (Saussure, Jacques Derrida); the humanities, Weatherby believes, “lost” language when it veered toward a post-structuralism obsessed with phonocentrism but lacking in the capacity to treat transformations of language in its non-semantic registers (like numbers and symbols). Chapter 2 builds the case that the sudden ubiquity of LLMs—language machines which speak back to us in legible sign systems—has forced the crisis through the realization that there is no cultural or humanistic apparatus with which to (quite literally) understand something like an LLM. Chapter 3 widens the aperture to the digital environment of the everyday, which Weatherby perceives as fundamentally language-driven (in modality, but also in the backend classification and generation systems which rely on linguistic operations). Chapter 4 unexpectedly invokes poetry and the literary: not poetry as in literal poems, but rather the language games poetry has always played as a kind of formal testbed for nonreferenetiality. Chapter 5, by Weatherby’s own account the culmination of the argument and the most challenging pages in the book, presents a “general poetics” of computational meaning. Here he articulates a deep structural homology between the Transformer architecture in machine learning (as introduced in 2017) and a semiological theory of language.

Weatherby’s core claim is that the one is not merely like the other but rather that both language and computation partake of common processes that are literalized and expressed for the first time in the language machines of his title and which now permeate the digital semiological surround. He locates the proof in the common currency of both systems: literal tokens, which is to say symbolic commodities, fungible and effectively indistinguishable deep within the vector architectures of the model. His final chapter and the conclusion each then gesture towards some implications of this thesis: language models (but also image models) as latent “ideologies,” and therefore a return to rhetoric as the disciplinary rubric to carry the whole project forward.

As the above suggests this is not a book to shy away from truth claims. Weatherby believes there are right and wrong ways of thinking (and talking) about language, cognition, and computation, and right and wrong ways of understanding what generative AI actually is and is doing in and through the world. Of course such a wide-ranging but relatively short book will not satisfy all comers: for example, it eschews any consideration of speech pathologies and aphasias, clinical phenomena that one might imagine could contribute something to the project of estranging language; likewise, the book seems to assume Indo-European languages when it speaks of language in toto, and its illustrations treat of only English and German.

Which brings us to culture, the other element of Weatherby’s forensics. Language models, he insists, offer something like an operative cultural map—not totalizing, but by definition indexical. Nor does computation stand outside of culture in this frame, always only ever relational to "culture." Like language, like math, like making pizza, computation is simply something humans do. Cultural AI is called out in the subtitle not as a plea to attend to AI’s better half but because culture has been detached from cognition for too long. If it’s language all the way down, it’s culture all the way home.