Language Without the World: Reading Thierry Poibeau’s Conversational AI
Thierry Poibeau’s Conversational AI is not a book that seeks to settle debates. It does not attempt to decide, once and for all, whether large language models (LLMs) think, understand, hallucinate, or merely simulate intelligence. This refusal is not a weakness but a methodological stance. Against the polarized rhetoric that dominates much contemporary discussion of generative AI—oscillating between technological enchantment and moral panic—Poibeau proposes something more demanding: a sustained inquiry into how these systems function, what kind of knowledge they mobilize, and what follows from their growing integration into social, intellectual, and political life.
At its core, the book advances a simple but consequential claim: LLMs should be understood less as cognitive entities than as complex, distributed artifacts, whose significance lies as much in their uses, institutional embedding, and effects as in their technical architecture. From this perspective, the central question is not whether machines “understand” language in a human sense, but what it means for language itself to be generated, circulated, and stabilized by statistical systems operating entirely within textual space.
Language Without Reference
Poibeau’s account of how LLMs work is deliberately non-mystifying. These systems do not perceive, act, or refer to the external world. They operate within language, modeling it through large-scale statistical regularities extracted from massive textual corpora. Words and phrases are encoded as vectors in high-dimensional spaces, transformed through layers of self-attention that allow the model to capture complex patterns of contextual association. The result is linguistic output that is often fluid, coherent, and pragmatically effective.
Yet fluency, the book insists, should not be mistaken for understanding. LLMs generate plausible continuations of discourse, not because they grasp meaning or intention, but because they have learned how language tends to behave under certain conditions. Their knowledge is correlational rather than referential; their competence lies in pattern completion rather than semantic grounding. Even when augmented with multimodal inputs, retrieval mechanisms, or reinforcement learning from human feedback, the fundamental epistemic structure remains unchanged: the system has no intrinsic access to truth, only to likelihood.
This insistence on language without world situates the book squarely within long-standing debates in the philosophy of language. Echoing critiques formulated by Bender, Koller, and others, Poibeau treats the “grounding problem” not as a temporary engineering challenge but as a structural feature of contemporary language models. At the same time, he resists the temptation to dismiss these systems as empty. Language, the book suggests, can remain functional, even socially consequential, without traditional forms of reference.
Writing as Palimpsest
One of the book’s most suggestive moves is to place LLMs in dialogue with structuralist and poststructuralist theories of text. Far from representing a radical break, generative models almost literalize ideas long articulated by Barthes and others: writing as recombination, authorship as distributed, originality as an effect rather than an essence. LLMs produce text as a palimpsest of prior discourse, an ongoing recomposition of citations, genres, and stylistic traces.
This has profound implications for contemporary writing practices. When a legal memo, an academic abstract, or a narrative paragraph is produced through interaction with a language model, authorship becomes blurred as it is distributed between user, model, training corpus, and institutional design. Poibeau is careful not to moralize this shift. Instead, he treats it as an invitation to rethink what writing means in an environment where machines can generate language without experience, presence, or subjectivity: “pure trace,” unmoored from voice.
For scholars attentive to the history of textual production, this analysis resonates strongly with concerns in the digital humanities. The book does not frame LLMs as neutral tools but as infrastructures that subtly reshape what counts as acceptable language, plausible argumentation, or stylistic norm.
Understanding, Cognition, and Anthropomorphism
A recurring theme throughout the book is the danger of anthropomorphism. Because LLMs speak fluently in the language of mind—offering explanations, judgments, even apparent self-reflection—they invite projections of agency, intention, and understanding. Poibeau systematically dismantles these projections. However impressive their performance, LLMs do not construct models of others’ beliefs, possess a first-person perspective, or exhibit temporal continuity of experience. They are stateless systems, generating responses token by token, without memory, motivation, or consciousness.
Recent techniques designed to scaffold reasoning, such as chain-of-thought prompting, constitutional principles, fine-tuning for explanation, may enhance usability and coherence, but they do not alter this basic architecture. The risk, Poibeau argues, lies not in what machines are, but in what we are tempted to see in them. LLMs function as mirrors of our interpretive habits, reflecting back our own assumptions about intelligence and language.
Creativity, Morality, and Authority
The book is particularly sharp in its treatment of creativity and moral judgment. Drawing on Margaret Boden’s typology, Poibeau acknowledges that LLMs can exhibit combinational and exploratory creativity, recombining familiar elements in novel ways and navigating structured conceptual spaces. What they cannot do is transform those spaces themselves. Lacking self-reflection, aesthetic judgment, and intentionality, machine-generated literature tends toward homogeneity rather than conceptual rupture.
A similar limit appears in the moral domain. When LLMs generate ethical judgments, they do so without deliberation or responsibility. Yet such judgments can acquire social force, precisely because they are expressed in authoritative, neutral-sounding language. The danger here is subtle but significant: persuasive fluency may be mistaken for moral understanding, leading to an uncritical delegation of judgment to statistical systems.
Power, Politics, and Synthetic Knowledge
Where Conversational AI becomes most urgent is in its analysis of power. Language models do not merely reflect the world; they inherit, reproduce, and often amplify the historical and social structures embedded in their training data. Bias, in this account, is not a technical glitch but a structural feature of systems shaped by institutional priorities, economic incentives, and platform dominance.
Poibeau extends this analysis to the production of synthetic knowledge. Techniques such as “silicon sampling,” which generate artificial populations for research or policy modeling, may offer practical benefits, but they also open the door to large-scale political manipulation. More broadly, the capacity of LLMs to generate persuasive text at scale transforms the dynamics of misinformation. Optimized for coherence rather than truth, these systems contribute to an environment in which synthetic narratives circulate faster than verification can keep pace, eroding epistemic trust.
The discussion of governance is correspondingly sober. Ethical guidelines alone are insufficient; regulatory frameworks matter. Yet governance, the book insists, is irreducibly political, shaped by struggles over economic power, technological sovereignty, and control of infrastructures. Platform lock-in, technofeudal dynamics, and cognitive capture are not incidental risks but structural tendencies of the current AI ecosystem.
Living With Language Models
Conversational AI does not conclude with a call to ban or embrace generative models. Instead, it advances a more demanding proposition: learning how to live with them. This means resisting ethical drift—the gradual delegation of judgment through habitual reliance—and insisting on keeping humans meaningfully in the loop, not merely as overseers but as moral and epistemic participants.
LLMs, Poibeau reminds us, are not moral agents. But they are objects of moral concern. They are made by human institutions, reflect human histories, and mediate human relations. To treat them otherwise is to misunderstand both machines and ourselves.
In this sense, Conversational AI is less a book about artificial intelligence than about language, authority, and responsibility in a world increasingly saturated with synthetic text. Its value lies not in offering definitive answers, but in providing a clear conceptual framework for asking better questions—precisely what is needed at a moment when fluent machines threaten to outpace critical reflection.
A.I. Use Statement
This blog post is based on my notes, which I asked ChatGPT 4.5 to analyze and cluster to reflect my reading and understanding of the book. This clustering served to organize an outline, which I used to write the blog post. The final styling and copy editing as done by ChatGPT 4.5, with final edits by myself.
OpenEdition suggests that you cite this post as follows:
Christian Henriot (December 23, 2025). Language Without the World: Reading Thierry Poibeau’s Conversational AI. Elites, Networks and Power in modern China. Retrieved June 18, 2026 from https://doi.org/10.58079/15evv
