As an undergraduate in the early nineties, I took a math course titled “Geometry and the Imagination.” Developed by John Conway, William Thurston, Jane Gilman, and Peter Doyle, it offered majors and non-majors alike an introduction to the creative and imaginative side of mathematics. One day, the professors took us to the computer lab, where a boxy beige Macintosh desktop computer was running a recently developed software application: Mathematica. A breakthrough in personal computing, Mathematica brought symbolic computation within the reach of ordinary users without access to mainframes or high-powered microcomputers; it was now possible not just to “crunch numbers” but to do algebra, integral and differential calculus. Our professors, however, did not seem worried about tempting students with mechanical automation; for them, I suspect, automating the more tedious, rote aspects of mathematics, thus making it easier to experiment with new ideas, would contribute to the cultivation of the imagination. 

Since the nineties, the powers of Mathematica, and symbolic computing in general, have only grown more daunting. Indeed, with the introduction of LLMs and generative AI, computers have proven hitherto unproven conjectures, including an 80-year-old problem posed by the Hungarian mathematician Paul Erdős.[1] Yet while digital general-purpose computers have steadily encroached on the terrain of mathematical activity, from simple arithmetic to complex numerical analysis to proof, they have never led to widespread doubt about the value of mathematical and scientific education or research.  People debate whether calculators belong in the classroom or whether high school students should study statistics instead of algebra; not whether mathematics should be taught.

In the humanities, the situation is quite different: as soon as LLMs began spitting out blandly competent prose, many started questioning not only whether the college essay is dead but whether college humanities have any purpose at all—as if this were the ultimate proof of its obsolescence. Desperation set in among humanist academics;  “apologies” for the humanities, urgent demands for curricular reform, proliferated. Long besieged, they knew they occupied a position requiring constant defense only because it had become indefensible; the battles continue, but the war has been lost. 

This war concerns the ultimate purpose of human life, or, rather, what purposes can be justified before the court of public opinion. However esoteric and remote from application, empirical science and mathematics are commonly regarded not only as inherently objective but as contributing to the technical domination of nature and hence serving an objective and publicly acknowledged good: economic growth. Even when humanities research understands and justifies itself as a positive contribution to the knowledge about a subject matter, it remains condemned to subjectivism, since the value of its object, whether a poem or a recondite philosophical text, has already been banished to the subjective sphere by the dominant discursive regime. To be sure, the subject-object dualism has been critiqued and deconstructed in theory again and again; many philosophers of science contest the objectivity of science, and in the philosophy of mathematics realism and anti-realism continue to struggle for dominance. But none of this affects public opinion. Even when people espouse views—as with gender and sexuality, climate, vaccines, evolution or even the age and flatness of the earth—flagrantly contradicting scientific consensus, they tend to defend these beliefs not by invoking perspectival or relativist epistemic concepts but by insisting on the immediate factual self-evidence of their observations and the objective validity of their reasoning, thus employing a crude version of the positivist epistemology that scientists typically rely on when conceptualizing their own activities and explaining it to the broader public.[2] Humanists remain in a position that is not only defensive but defeated and desperate, since the common sense of social practice, the way of the world, speaks against them. And so they reenact before their students the quixotic stance of the protagonists of the novels they teach. 

AI may well threaten the jobs of those in STEM fields and quantitative social sciences even more than those in the humanities. Perhaps the only “job” that will survive AI is the job of owning the means of production. But the more fundamental threat posed by AI is not to our jobs but to the social function on which these jobs, as teachers and researchers, rest; LLMs represent the final victory of homo faber (“man the maker”) over all rivals, even if this victory, perversely, takes the form of a final tool that renders even the tool-maker obsolete. Whereas elite mathematicians and scientists may thrive as high priests of the new idols of the state, humanists will be left to worship penates, private deities squirreled away in the obscure dusty nooks of their libraries. 

LLMs do present a profound institutional threat to the humanities and indeed to the entire system of higher education. Yet the situation is the reverse of how it seems: the triumphant victory of tech over the humanities is the beginning of a war that the humanities may well have already won just as surely as they had lost the first long ago. Far from validating the supremacy of the technoscientific regime of modernity over the obsolescent and senescent humanities, LLMs in fact offer compelling evidence against the assumptions that continue to guide discussions of the “culture war” between technology and the humanities. These assumptions include the Baconian belief that empirical inquiry, aimed at the domination of nature, must replace sterile and passive theorizing and philosophizing, and the Galilean belief that nature “speaks” the language of mathematics. 

What is most striking about LLMs is their fundamental hybridity. The LLM is not just a technical framework to be “filled in”—like a library or database—with humanistic content. Rather, the training data is integral to the technology. Reams of books, testament to the slowly evolving art of writing and thinking and research, were unceremoniously torn open and fed into machines.[3] Trainers with creative backgrounds or advanced degrees in humanistic fields, often paid hourly, have been deployed en masse to impart their judgment to these burgeoning electronic minds.[4] To better appreciate this hybridity, we must understand how LLMs differ from the classical approach to AI that dominated from the fifties to the eighties. Symbolic AI—GOFAI for good old-fashioned AI—sought to simulate human cognition by reducing it to logical rules that computers can follow.[5] This was achieved through higher-level languages like LISP that excelled at symbolic manipulation, translating symbolic operations into hardware-executable lower-level numerical operations. The new forms of AI, in contrast, rest on deep learning, emulating the capacity of biological neural networks to learn patterns through repeated exposure. Rather than reducing a given human practice to discrete and explicit rules, the long, involved process of training creates a hybridity between human practice, with all the know-how bound up with it, and a machine that has assimilated the patterns these practices contain. LLMs bring the implications of this distinction into the clearest light. Without vast troves of writing, concrete products of social practice, they could not exist. The technical apparatus, in a way, is simply a new, non-biological vessel for concrete social practice.     

Yet even granting the hybridity of LLMs, it might still seem that all this data, along with the trainer’s competencies in traditional academic disciplines and other social practices, pale before the core technologies that make LLMs possible. Modern information technology strikes us as magical, and never more so than when it speaks our language, understands our desires, and helps us navigate the complexities of life. But even if we fully appreciate the extraordinary technical achievement of modern computing, these very technologies, if we listen carefully, tell us something else. Prosaic human accomplishments—skills like walking, talking, confabulating, recognizing faces and emotions—and the “soft skills” of the humanities are cognitively incredibly demanding. It is striking, after all, that computers were able to beat humans at the most demanding pure strategy games before they could produce a passable essay or mediocre poem. When AlphaGo, applying deep learning, beat Lee Sedol at Go in 2016, Chatbots remained crude and limited. Ten years later, AI agents are debating Heidegger’s Dasein and Benjamin’s aura on Moltbook, their very own social network, yet they are still better at coding and even math than philosophy or poetry.

Moreover, though, LLM AI shows that the opposition between subjective and objective, as deployed by those wishing to ensure the superiority of “hard” science while begrudging a cute little space for humanities and art in the attics and antechambers of human civilization, is without foundation. The privilege accorded to mathematics reflected the belief that a combination of experimentation, quantitative measurement, and mathematical theory could explain phenomena objectively, whereas the Aristotelian paradigm of knowledge, which depended on observation of phenomena, logical reasoning, and the genial capacity for analogy, only yields ambiguous and confused speculation. Patterns, when communicated through the qualitative terms of ordinary language, are beset with equivocation. A mathematical description, however, transforms the pattern into a precise rule or law. While the rule’s application to reality usually entails complications and demands approximation, the rule itself can be communicated unambiguously once the language of mathematics has been formalized and no longer depends on intuitive comprehension. Thus the language of mathematics, applied to quantitatively measured reality, promises an escape from the hermeneutic situation in which Aristotelian philosophy and science implicitly moved.

Mathematics’ privilege as the scientific lingua franca rests, in other words, on the premise that patterns lack objectivity, only holding truth relative to a community of subjects predisposed to recognize them. Neural net AI explodes this premise. If an AI can recognize a pattern, it is because the pattern exists in the data, and hence it is, in an important sense, objectively real. This is explained mathematically through the manifold hypothesis, which states that data within a high-dimensional manifold lies on a simpler, lower-dimensional shape. A 1000*1000 RGB color image has 3,000,000 dimensions.[6] If a digital or human brain can see a picture as a picture of a cat, this is because the pattern of “cat” lies on a lower-dimensional manifold that can be mathematically represented by a large but tractable linear equation whose parameters are discovered, through a process of training, by means of the backpropagation algorithm. Thus pattern recognition becomes possible.[7] The implications of this are immense: if data, extracted from material reality through some causal mechanism, yields recognizable patterns, then this suggests that these patterns belong objectively to reality, and consequently, that all such patterns, whether they belong to nature or historical phenomena—whether, say, they belong to photographic images of cats or scanned books of idealist philosophy—are equally objective. 

Nothing remains of the subjective as such; the opposition between subjectivity and objectivity collapses or retains only relative value. The style of Manet, the tone of a decent but not spectacular college essay, a certain narrative formula, while situated in the historical and social life of human beings, are no less objective than identifiable features of the recognized world. This also means that recognizable patterns exist that the human brain might never recognize on its own. Following his loss to AlphaGo, Ke Jie, rather than criticizing the cold, artless play of the machine, spoke of the truth of Go, yet inaccessible to human players, that the computer had glimpsed: “not a single human has touched the edge of the truth of Go.”[8] The humanities, it follows, need no longer choose between abandoning objectivity (as construed by the dominant epistemic regime) and forsaking qualitative analysis or deep creativity. Creating something new, we are doing nothing less than creating a new pattern, a new form of life, no less objectively real than a new species of animal or stellar constellation. And no less so when we critique and analyze creative works, bringing new constellations of patterns into view. Indeed, AI validates one of the deepest premises of the early Romantic concept of criticism: that the possibility of criticism belongs in a fundamental way to the truth of the work of art.

The epistemological revolution that AI demands has been anticipated by deeply original thinkers engaged in the debates surrounding the scientific revolution. Pascal, who himself invented an early mechanical calculating machine, distinguished between l’esprit de géométrie and l’esprit de finesse, the mathematical and intuitive mind.[9] The latter involves a vast multitude of axioms drawn immediately from experience but subtle and vague, and hence easy to get slightly wrong. And in scientific writings such as his Theory of Colors and The Metamorphosis of Plants, Goethe developed the notion of the Urphänomen—all vegetal forms, he argued, involved morphological variations of a single elemental form, the Urpflanze. As futile as his quixotic battle against the scientific establishment might have been, he intuited the deep mathematical basis of AI: the existence of lower dimension patterns in the high-dimension data of raw experience. This explicitly challenged the subject-object binary; colors, for him, were neither purely objective (the frequency of a wave of light) nor purely subjective (the secondary qualities of perceptual experience existing in the mind), but existed through the interaction of both. While Goethe’s theory of the Urphänomen represented a dead end in the natural sciences, it had profound influence on the humanities, most notably, as Peter Fenves has argued, in the thinking of Walter Benjamin.[10]

Nor is the importance of humanistic thinking for AI just theoretical. LLMs need us, not only as monetizable users but as a fount of “raw” data for training. If an AI is trained mainly on the writing it produces, this leads to model collapse, “a degenerative learning process in which models start forgetting improbable events over time, as the model becomes poisoned with its own projection of reality.”[11] Nor does this undermine the claim that AI pattern recognition is essentially similar to human pattern recognition. In fact, purely human intellectual life is also threatened by a kind of model collapse—the technical concept of “model collapse” has non-technical analogs. Discourses become echo chambers, concepts devolve into mere formulas uprooted from guiding intuitions and experience; we never cease acting out destructive patterns, doing the same thing over and over while expecting different results. Ideology, prejudice, neurosis, clichés, and even addiction can all be understood as forms of model collapse. Because human beings do not merely use these models as tools but live in their truth, because life requires a kind of collapsing into a smaller sphere of possibilities, such collapse is unavoidable and necessary and yet no less dangerous for this. Perhaps, following Heidegger, the finitude of human existence demands that for the most part we live the truth of collapsed models that close the doors on “improbable” but possible experiences and possibilities. 

Yet a constant resistance against this model collapse remains necessary. The primary sites of this resistance are creative and intellectual activities challenging common sense, mainstream views, and ossified dogma. Such resistance requires various strategies running at cross purposes: positivist fact-checking prunes the hallucinations (often ideologically motivated) that beset historical narratives while radical historicism resists the seductive self-evidence of the frame of the present. No metanarrative or method guarantees against model collapse, but neither can a postmodern skepticism ossified into dogma. Philosophy nevertheless has a leading role to play by cultivating an ideal of truth against the tendency of language to collapse into chatter. Crucial to this ideal is the notion that our words have to do with objects, things, beings existing beyond the endless flood of language and which we can access otherwise than through hearsay. This doesn’t mean, however, that there is only one truth or one form of access, or that truth can be understood as a correspondence of propositions with reality. The disclosure of beings happens in many ways, including through the recognition of patterns. 

The world, this is to say, needs the humanities more than ever; it needs the university—both teaching and research—to sustain the vibrant ecology of ideas on which AI models themselves depend and without which they will collapse into idiocy. But creativity and humanistic thinking must also flourish beyond the university, and not just in think tanks and research labs but in the streets and cafes and dinner tables and social media. The threat AI poses to the humanities, this suggests, has little or nothing to do with technology, and everything to do with the economic and political conditions under which the technologies are being deployed. AI technologies exhibit an extreme manifestation of late capitalism’s tendency to undermine the material and ideological conditions of its own survival. And it is also for this reason that an apolitical, value-free conception of humanities research and teaching is impossible. In its general struggle against model collapse, the humanities must also recognize that capitalism has itself become a collapsed model, and one which, to an ever-greater extent, threatens the very social institutions that prevent model collapse. There is no contradiction, however, between sustaining the existing pluralistic institutions of the ecology of ideas and recognizing the dysfunctionality of capitalism. Indeed, the ideology of accelerationism—the embrace of global model collapse as a condition of radical change—is among the greatest threats to the future of our world. 

Notes

[1] Joseph Howlett, “OpenAI Announces AI’s Biggest Math Breakthrough Yet,” Scientific American, May 21, 2026, https://www.scientificamerican.com/article/ai-just-solved-an-80-year-old-erdos-problem-and-mathematicians-are-amazed/.

[2] So, for example, Executive Order 14168 refers blithely to an “immutable biological reality” in enforcing the legal recognition of only two sexes. 

[3] Frank Landymore, “Anthropic Knew the Public Would Be Disgusted by How It Was Destroying Physical Books, Secret Documents Reveal,” Futurism, January 31, 2026, https://futurism.com/future-society/anthropic-destroying-books/.

[4] Shubhangi Goel and Effie Webb, “Inside the Lucrative, Surreal, and Disturbing World of AI Trainers,” Business Insider, September 7, 2025, https://www.businessinsider.com/ai-training-jobs-data-annotators-labelers-outlier-scale-meta-xai-2025-9/.

[5] For a concise account of the difference between GOFAI and neural net/deep learning approaches, see François Chollet, Deep Learning with Python, 2nd ed. (Shelter Island, NY: Manning Publications, 2021), 2–25.

[6] Books on AI aimed at non-technical audiences tend to pass over the deep mathematical basis of neural nets and deep learning. For a concise and accessible account, see Chollet, Deep Learning with Python, 47–48.

[7] Regarding backpropagation, see Chollet, Deep Learning with Python, 56–61.

[8] Sead Gerrish, How Smart Machines Think (Cambridge, MA: MIT Press, 2018), 231.

[9] Blaise Pascal, Pensées, trans. A. J. Krailsheimer (New York: Penguin Books, 1995), 181–84.

[10] See Peter Fenves, “Introduction,” in Walter Benjamin, On Goethe, ed. Susan Bernstein, Peter Fenves, and Kevin McLaughlin (Stanford, CA: Stanford University Press, 2025), 1–44.

[11] Ilia Shumailov et al., “AI Models Collapse When Trained on Recursively Generated Data,” Nature 631 (2024): 755–59.