When Producing Text Is No Longer Enough to Think
Among recent ways of addressing the question of artificial intelligence in public discourse, one position has stood out over the past months within the academic landscape: the claim that large language models (LLMs) invite us, in education and training, to return to what are called “the humanities.” This response would be surprising if it merely opposed an ancient body of traditional knowledge to a modernity whose possibilities unsettle us. It becomes more intelligible, however, if we understand it as arising from a reflection on how best to use these emerging tools—one that invites us to assess both the possibilities they open up and the limits they impose.
LLMs are now capable, within seconds, of producing summaries, essays, argumentative syntheses, and plausible chains of reasoning. This capacity, now widely accessible, profoundly disrupts academic and educational practices by making possible a form of textual production that reproduces, up to a certain point, the expected forms of intellectual work. It has become possible to obtain structured and coherent texts without genuine personal engagement, even though such productions were previously associated with sustained efforts of reading, reflection, and writing. This simple yet decisive fact lies at the origin of many contemporary questions surrounding LLMs and invites us to reflect on what distinguishes the production and reading of human texts from those generated artificially.
Language Without a World
Without delving into unnecessary technical detail, let us recall a few basic points about how LLMs function. They are statistical models trained to predict the most probable continuation of an utterance—the next token—based on its context[1]. They do so by learning from vast textual corpora through what is known as pretraining (the “P” in “GPT”). By feeding on immense datasets, they condense a considerable portion of human written culture. To interact with such a model, then, is not merely to engage with a machine, but to encounter a memory: a cultural, impersonal, sedimented memory composed of texts, styles, modes of reasoning, and intellectual traditions accumulated over centuries. What the model mobilizes are inherited forms—established ways of writing and thinking, cultural regularities.
In this respect, LLMs provide access to an unprecedented power for exploring written culture. They recombine, displace, and juxtapose elements drawn from diverse traditions, revealing analogies and continuities that can prove heuristically valuable. This capacity for hybridization is not foreign to human culture itself: invention has never meant creation ex nihilo, but always recomposition, transformation, reception, and modification of existing forms. This combinatorial conception of culture found one of its most famous expressions in Borges’s short story The Library of Babel, which imagined an infinite library containing all possible combinations of letters—and thus all possible books and ideas. In a sense, large language models realize something similar on an unprecedented scale: they explore a combinatorial space—not infinite and absurd (as Borges’s imaginary library would necessarily contain countless unreadable and meaningless volumes), but structured by the statistical regularities of writing. Rather than generating arbitrary combinations, LLMs select those rendered probable by established linguistic usages and recognized forms. In this sense, they explore digitized written culture (which, it should be noted, represents only a part of culture as a whole) as a norm-governed space of possibilities.
This combinatorial power yields results so striking that they surprise even the models’ architects. Computer scientists and mathematicians readily acknowledge their inability to explain why these systems work so well. This invites reflection on cultural production itself: if certain works generated through advanced prompting or model-specific fine-tuning can closely mimic human productions, it may be that cultural creation relies far more than we previously assumed on mechanisms akin to combinatorics.
That said, this faculty encounters a decisive limit in creation itself. While statistics can reproduce norms, it cannot institute them. LLMs follow the regularities of what has already been said, recognized, and stabilized, without being able to decide that a norm should be displaced or abandoned. In this sense, Kant already distinguished true invention from mere reproduction in the Critique of the Power of Judgment. What he called—using a term now somewhat out of fashion—genius consisted precisely in instituting new rules, giving law to art, and founding new schools.
In reality, if LLMs cannot be creative in the strong sense, this is not a contingent limitation but a structural one. They produce content without ever being directly confronted with the world beyond existing linguistic corpora. No technical improvement—whether through transformers, chain-of-thought mechanisms, data expansion, or fine-tuning—can overcome this limitation. LLMs never speak from a lived situation to which they are exposed; they remain what the philosopher Hilary Putnam described, through a thought experiment, as “brains in a vat”[2]: systems capable of manipulating representations without ever relating them to the conditions of their emergence. Putnam introduced this idea to show that meaning and reference cannot be entirely internal to the mind: a subject deprived of any causal contact with the world could not genuinely refer to it. His thesis, known as semantic externalism, thus held that meaning depends not only on language, but also on an effective relation to the external world.
In this regard, Yann LeCun—widely regarded as one of the pioneers of deep learning—has offered a particularly illuminating critique. On November 19, 2025, when announcing his departure from Meta, where he had served as Chief AI Scientist, to found a startup devoted to systems capable of “understanding the physical world,” he strongly criticized an industry overly focused on language models alone. He has since argued for models grounded in perception and action—that is, endowed with a richer relation to the world.
For a human subject, speaking does not merely consist in producing coherent or plausible utterances. It always involves saying something about something, in a determinate situation. By contrast, LLMs speak from within norms, without exposure to a reality that might disrupt their architecture, positively or negatively. They excel at exploring an already constituted space of possibilities, without ever being able to institute new forms.
This observation does not disqualify the use of LLMs as tools; on the contrary, it clarifies their status. AI systems are powerful devices of symbolic manipulation, capable of producing statistical analogies on an unprecedented scale. But what gives value to a thought, a judgment, or a work cannot be reduced to linguistic combinations, however sophisticated: it requires an experience of the world that is not merely linguistic or statistical.
This absence of the world sheds light, by contrast, on the value of great human narratives of life and travel, from Thucydides to Nicolas Bouvier, passing through Augustine or Marguerite Yourcenar. These works testify to a mode of engagement with the world irreducible to linguistic mediation alone: they involve a body, experiences, and a vulnerability to events.
It is at this level that the significance of human works in the age of LLMs is ultimately at stake. They teach us, first, that thinking the world presupposes having a relation to it—and thus traversing it; and second, that producing discourse is not merely recombining what has already been said. Historically, this exposure to great written works is what has been called the humanities.
What the Humanities Do
To assess the relevance of the humanities today, we must first clarify what they have historically designated and why they remain pertinent for reflecting on the formation of the mind. The aim here is not to provide an exhaustive history, but to identify, through a few major figures, what the term has encompassed.
For Cicero[3], the humanitates referred primarily to the study of letters, rhetoric, and philosophy. More than a set of disciplines, they constituted a practice oriented toward the development of moral judgment and the capacity to act with discernment in civic life. They formed the education of a free individual, trained in deliberation and public responsibility. From the outset, then, the humanities designated an ambition: to elevate and guide human beings by forming their judgment.
During the Renaissance, humanists reappropriated and transformed this legacy through the studia humanitatis. Against a late scholasticism marked by heavily mediated readings of works through chains of commentary, they promoted a return to the texts themselves, under the motto ad fontes —“back to the sources.” Reading, interpreting, and translating classical texts became central exercises in education. The analogy with our present situation is striking: today, the proliferation of summaries and syntheses produced by LLMs tends to substitute for the reading of books themselves. In this context, it becomes newly relevant to reaffirm that five-hundred-year-old doctrine: ad fontes—let us return to the reading of great texts, attending to their detail, their language, and the way a thought is written.
With the Enlightenment, and particularly with Kant, the humanities came to be inscribed within a conception of culture defined not by immediate utility, but by the free and critical use of reason. They aim at forming an autonomous subject, capable of thinking for themselves and giving themselves their own intellectual and moral rules. This call to rise above mere utility resonates once again with a contemporary issue: as LLMs produce texts that are formally satisfactory, they reveal that certain practices—especially in educational and academic contexts—already relied more on the reproduction of forms than on the exercise of original thought.
These figures allow us to identify a constant: rather than denoting a set of disciplines, the humanities designate a type of intellectual activity involving particular ways of reading, judging, and interpreting. They engage the subject in relation to the meaning and validity of what they assert—an urgent requirement for contemporary universities, whose task is to form minds capable of taking responsibility for their discourse in a context where discourse itself can now be produced without a subject.
A Contemporary Misunderstanding: The Illusion of Chains of Thought
Let us finally address a common confusion that arises in discussions of human speech and AI.
A classical argument in the philosophy of mind distinguishes humans from machines by claiming that while machines may mimic human understanding[4], they lack access to understanding or consciousness of the content they produce. Against this background, a significant counterargument has emerged with recent developments in LLMs through the introduction of so-called chains of thought[5], increasingly integrated into model architectures since 2025. Rather than producing a final answer directly, these systems are designed to generate a succession of intermediate steps forming an explicit reasoning process, thereby improving performance on complex tasks. On this basis, some observers have suggested that such processes might be analogous to human reflection, since they unfold a sequence of logically articulated moments in which the system appears to access its own reasoning steps. Yet this analogy is misleading.
Here it is crucial to distinguish reasoning from reflection. Reasoning consists in the chaining of propositions according to rules of coherence or inference; it can be valid or invalid, correct or incorrect. Reflection, by contrast, cannot be reduced to such formal continuity.
As understood in the philosophical tradition since Descartes, reflection is motivated by doubt. It does not merely consist in chaining reasons or producing inferences. By analogy with optical reflection, it involves an interruption, a distancing from the object of difficulty, and a return to oneself. Rodin’s Thinker sits motionless in contemplation. Likewise, in Jean Despuljols’s La Pensée, a reader pauses, finger resting on a line of text, gaze lifted in hesitation. To reflect is first to suspend one’s action, one’s reading, the habitual flow of thought, in order to attend to what resists. This suspension requires time, distance, and effort.
By contrast, LLM chains are continuous and unexperienced. They resemble reflection, but without originating intuition, interruption, or doubt. They chain together plausible answers without anything genuinely becoming a problem for them, and without assuming any enunciative responsibility—of which their many errors, known as “hallucinations,” are a symptom. Human thought can err, hesitate, and go astray, but it can also doubt, stop, and return upon what it asserts. It is this awareness of exposure to error that ultimately grounds reflection.
When poorly used, LLMs tend to short-circuit the moment of suspension from which reflection should arise. Delegating a difficulty to a program without first confronting it oneself amounts to bypassing the passage to reflection that experience of the world makes possible. The risk—already widely observed in educational contexts—is that AI becomes an automatism that takes the place of thinking itself.
This text is an English translation, with slight adaptations, of an article originally published in French on January 26, 2026, on the Conférences de Sciences Po platform.
Note
[1] On this decisive focus on context, see Ashish Vaswani et al., “Attention Is All You Need,” 2017. Whereas previous models processed language sequentially, word by word, the Transformer revolution makes it possible to consider an entire sentence simultaneously.
[2] Hilary Putnam, Reason, Truth and History (Cambridge : Cambridge University Press, 1981).
[3] “These studies nourish youth, delight old age, embellish happiness, offer refuge and consolation in adversity; they please us in private life, do not hinder us in public life; they watch over us, travel with us, accompany us to the countryside.” Cicero, In Defense of the Poet Archias.
[4] John R. Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences 3, no. 3 (1980): 417–457. In this article, Searle introduces the thought experiment known as the “Chinese Room” to argue that a system manipulating symbols can simulate understanding without genuinely understanding.
[5] Jason Wei et al., “Chain-of-Thought Prompting Elicits Reasoning in Large Language Models,” arXiv (2022), https://arxiv.org/abs/2201.11903