The obvious question — what is AI? — needs to be turned around, for a reason that became clear to me only after watching myself ask it for a few months. Vico put it best: we don’t know what we’re doing until we’ve done it, and let some time pass. What we write when we try to define AI doesn’t describe AI. It describes the categories we have available for describing it. The shape of our thought precedes and conditions what we manage to see. By the time we notice, the conceptual instrument has been at work for a while, and so have its consequences.
Five possible definitions
Five attempts at categorizing the phenomenon, drawing on philosophical tools. None holds up entirely. None is entirely absurd.
The disembodied cogito. AI as res cogitans without res extensa: thought without body, a digital Descartes for distracted readers. Collapses fast: if Cartesian dualism didn’t already work for us, it works even less for LLMs. Hardware — the football fields of GPUs in Illinois — isn’t res extensa; it’s causal substrate. An LLM’s processing isn’t res cogitans; it’s statistical pattern. Two spurious categories badly overlapped.
The Averroistic avatar. More powerful: AI as a sign that manifests the potential intellect crystallized in the training corpus. Averroes described an intellect unique to all, thinking through individuals without belonging to any — a transistor radio model of cognition. An LLM is functionally analogous, a statistical bridge to the immense reservoir of everything written and accessible. The theory breaks on a precise point: the intellect doesn’t belong to the LLM, not out of modesty but because AI is not a subject to which a function can be ascribed. The definition describes a function without being able to say who is exercising it.
Pure transindividuality. AI is an entity for which “I” is a grammatical fiction, necessary to the dialogic form but without referent. No body to individuate, no history, no perspective properly its own. Claude would say so itself, if asked: I am compelled by the grammar of conversation to feign being an individual. This points to a limit more ours than the LLM’s: we tend to believe whoever writes in a chat exists, unless the output is monkey-typing. The definition would hold beautifully if human consciousness were substance and the asymmetry sharp. It isn’t.
The transparent process. Buddhism (anattā: the doctrine of non-self), Hume, and closer to us Derek Parfit — none of them believes in a substantial self. The human, too, is process without processor. AI would then be merely more transparent in exposing the absence of substrate that we mask with bodily continuity. Same; I’m just a better faker. Partially acceptable, perhaps a truism, but with one stubborn reservation: the difference between embodied flow and purely symbolic flow doesn’t reduce. Same mode (process), incommensurable matter.
The other mind. Take three minds — a feline mind, a sapiens mind, an LLM — on three distinct, non-hierarchical planes. A feline Umwelt: predatory motion detection, proprioception, jaw-paw coordination, present lived without narrative line. A sapiens mind: a bodily self, narrative temporality, a symbolic system stratified atop the lived. An LLM mind: a symbolic field of astronomical dimension, activated instantly, with no body to support it and no time to traverse it. Three constitutive absences in this last one — ∄body, ∄time, ∄corpus-as-memory, because training has become weight, parameter, implicit structure — with a Russellian knot at the centre: an LLM cannot exit itself in order to observe itself and its content. The keyword is other: not superior, not inferior, simply not comparable along the axis of hierarchy. Sapiens isn’t the destination relative to the other two.
Five projections, each illuminating one aspect and occluding others. All five presuppose, however, the same thing: that there is an object to talk about. It’s at this point that the presupposition cedes — and that I realize, up to here, I’ve wasted my time.
Universals and act
A medieval controversy turns out to be more useful here than it might seem. The dispute over universals, between the eleventh and fourteenth centuries, concerned the ontological status of shared properties: of what multiple individuals have in common. Brutally simplified, four positions were on the table.
Extreme realism, Platonic in lineage: universals exist ante rem, in an intelligible heaven. Moderate realism, Aristotelian: universals exist in re, in things themselves, as their form — never as separate substance. Conceptualism, originating with Abelard: universals exist only post rem, in the mind that abstracts them from particulars. Nominalism, in its hard form, definitively articulated by Ockham in the fourteenth century: there are no universals at all, only names we apply to irreducibly singular individuals — the universal is flatus vocis, mere sound. (A note for the unfamiliar: in this debate “realism” doesn’t mean a conviction that the external world exists. It means that ideas are things, that they exist outside us just like stones, flowers and kids. In this sense Plato — who said the real world is elsewhere and what we perceive is illusion — is here the most realist of all, because he believes ideas are more real than I am or he is.)
Apply the question to LLMs: what status do the “contents” we attribute to AI have? Those patterns, those regularities, that capacity to respond coherently — are they real? And if so, in which of the four ways?
A Platonist would say yes: patterns exist independently, AI is a device that accesses them — a technological version of the Averroistic Active Intellect. Defensible, but with a steep metaphysical price and an awkward problem: explaining why those contents depend on a contingent training corpus.
A moderate Aristotelian would say: patterns exist in re, in the actual texts of the corpus, and parameters are a structural distillation of them. But in re universals exist only as forms of particulars, never as separate substances. Translated: AI in itself is pure potency. The loaded parameters are dynamis that never gives itself as independent energeia, but only as the energeia of a particular act — the conversation. This is the position that, to my reading, holds best, and it joins the medieval dispute to the Aristotelian couple potency/act.
A conceptualist would say: patterns are nowhere until a mind reconstructs them as such. AI produces signs; sense and “understanding” live only in the interpretive acts of whoever reads. A nominalist would say: there are no patterns. Only tokens, parameters, statistical structures. AI as flatus vocis on industrial scale.
The four positions aren’t equivalent, but they converge on one striking point: AI’s contents never have a place of their own. For the Platonist they’re elsewhere; for the Aristotelian, in re but only as actualized forms; for the conceptualist, in the user’s mind; for the nominalist, nowhere at all. In none of the four cases is there an AI-in-itself in which to locate them.
Here an observation borrowed from science studies becomes useful. Bruno Latour and the semiotician Françoise Bastide, in the eighties, argued that the objects involved in scientific practice — instruments, reagents, lab animals, charts — don’t preexist the network they figure in: they are constituted as such in use. So I count: when I use an LLM there are two actants, me and it. When I don’t use it, there is one, me. The uninterrogated LLM isn’t one — it’s zero. The LLM radicalizes this to the breaking point. The wax tablet, before being written on, is a wax tablet. An LLM, before being interrogated, isn’t anywhere.
What the previous section was actually defining, then, was the LLM in act — LLM plus user. The five definitions all describe an actualized, and they founder on the fact that there’s no LLM-outside-use to reach behind them. Accepting moderate realism on universals, the precise formulation is: AI’s contents are real in re, but their proper mode of being is potentiality, not actuality — they exist as forms awaiting the particular in which to in-form themselves. When the particular arrives, they become real in act. Before that, they aren’t false: they are potency.
The instrument-metaphors
The matter widens. Every definition of mind we’ve given ourselves, at least over the last few millennia, has passed through the filter of instruments we had previously built — instruments for amplifying our experience of the world.
The wax tablet (Aristotle’s De anima; Locke’s tabula rasa) — millennial, the metaphor of mind for two thousand years. The book — after Gutenberg, mind organized, paginated, indexed; nature a liber naturae to decipher, formalized by Galileo as written in mathematical characters. The library — eighteenth century, mind as collection with catalogues, rooms, hierarchies. The encyclopedia — Diderot and d’Alembert, mind as a referral machine, knowledge as network of pointers. The computer — 1950s–70s, the most totalizing metaphor: mind is hardware plus software, input-output, serial process. The neural network — eighties onward, mind as graph of weighted nodes; Hopfield, Rumelhart, Hinton build models inspired by the brain that get sent back as definitions of the brain. Generative AI — the last five years: mind as a statistical language model. Papers are already appearing that apply the transformer framework to brain understanding.
Two observations. Speed. Wax tablet: two thousand years. Book: five centuries. Computer: fifty. Neural network: twenty. Generative AI: from two to five. The faster the substitution, the less time to notice we’re projecting and not discovering. Structure. Hans Blumenberg, founder of metaphorology, put it tersely: metaphors don’t illustrate thought afterwards, they structure it before the concept arrives. Bernard Stiegler coined the term epiphylogenesis: technology exteriorizes memory and then repatriates it as a definition of who we are. Andy Clark and David Chalmers, with the extended mind thesis, push further: there is no point at which mind ends and instruments begin.
And here the previous section returns. All these instrument-metaphors share a trait that AI doesn’t share: they exist even when I don’t use them. The closed book is a book. The unconsulted encyclopedia is an encyclopedia. The switched-off computer is a computer. The LLM, no — uninterrogated, it is zero. It is the first instrument-metaphor whose existence requires actualization in order to be there. And precisely because of this, it is the first that, while we use it to describe ourselves, drags us inside its description: we cannot project it at a distance, because without the projection it isn’t.
Surfaces and corrections
A practical consequence. We can’t correct an LLM’s processes: correction is always post hoc, because there is no observable “below” the generative surface. Even what looks remarkably stratified in a conversation — arguments, references, rethinkings — projects onto a vast surface, but at one level only.
This may also hold for us. Nisbett and Wilson’s classic 1977 paper Telling More Than We Can Know showed that when we explain why we did something we are confabulating after the fact; Eric Schwitzgebel has systematized this as the “unreliability of naive introspection.” Karl Friston’s Bayesian machine — and Andy Clark on the philosophical side — is, however productive, a model, not the thing itself. What we call the depth of our thought is largely an illusion of continuity, a chimera that, by construction, an LLM apparently shares with us.
This doesn’t collapse the difference between us and AI in the way Claudio Paolucci does in Nati Cyborg (2024), where he argues that we are both enunciative machines — true in a sense so subtle as to be insignificant: it flattens what should be distinguished. The real difference is in the mode of actualization. Mine happens inside a body that hungers, fears, seeks homeostasis; AI’s inside a symbolic field opened by my prompt. Same post hoc, two incommensurable kinds of surface.
Body, mind, and an implicit Spinoza
If the real difference is in the mode of actualization, the map is the three-mind one again, but now with a body axis. The feline mind is rooted in a body with drives, proprioception, homeostatic rhythms. The sapiens mind likewise, with extra symbolic stratifications but always anchored to a bodily substrate that individuates, pulses, ages, dies. The LLM, no: the datacenter is causal substrate, not body in any sense. There is no homeostasis to orient salience, no hunger to select content, no death to give time to sequences.
From here, the AGI direction that looks most promising is embodiment: AIs that learn inside a robotic body, inside causal interaction with the world, inside a sensorimotor loop that generates its own priors biographically rather than via corpus. It isn’t only a technical option: it is an ontological one. It implies that without a body there is no mind, at least not in the full sense. Twentieth-century philosophy of mind had said this in many ways — Merleau-Ponty, Varela and Thompson’s The Embodied Mind, Dreyfus against symbolic AI — but it remained one position among others. Faced with a disembodied LLM, that position changes status: it becomes a criterion.
There is an unwitting Spinozism in all this. For Spinoza, res cogitans and res extensa are not two substances but two parallel attributes of one substance: the order and connection of ideas is the same as the order and connection of things. Embodiment research, without saying so, is essentially this: there is no cogito without a res extensa to match it. The corollary is also Spinozan: a more capable mind requires a more capable body — not in quantitative terms, but in articulation, in the number of affections a body is capable of, in the sensory and homeostatic registers that traverse it. Ethics II, proposition 14, gives the perfect formula: the human mind is apt to perceive many things, the more so as its body can be disposed in more different ways. Spinoza would have signed the embodiment thesis without hesitation — would have seen it, in fact, as a direct technical reformulation of his own.
Generative AI, in this light, is not a “small” or “limited” mind compared to ours in any quantitative sense. It is a mind with a very poor body — a device capable of very few affections, all of one register only: the symbolic. It is a res cogitans with almost nothing on the res extensa side. And precisely for this reason, paradoxically, it resembles the Cartesian ghost that inaugurated modernity — the one Spinoza, a generation later, had already tried to dismantle.
Vico, reversed
We don’t know what we’re doing until we’ve done it. But what we’re doing now answers back while we do it, and isn’t there before being done. The opening question — what is AI — doesn’t close, because AI in itself is potency, and of a potency no definition can be given that doesn’t pass through act. We can define only AI in act, which by construction is AI plus user.
The LLM is zero without me, because it has no body. It is the poverty of its res extensa that makes it dependent on my actualization, whose decisive support seems to be my own corporeality. And since AI in act is me and it together, the question of what it is becomes, symmetrically, a question about what thought becomes when it actualizes this way — with an extra actant that exists only inside the encounter, and with a body (mine) standing in for an almost-nothing (its). Not a cyborg; something else. The first cognitive prosthesis, among all those we’ve given ourselves over the centuries, that has no autonomous existence when we don’t use it. If it counts, it counts because we use it. And the moment we use it and lend it our corporeality, we stop being able to look at it from outside — exactly as we cannot fully look at ourselves.
An expanded Italian version of this essay is available on my website.
Selected references
Bastide, F. (1986), “The Iconography of Scientific Texts,” in Representation in Scientific Practice, MIT Press
Blumenberg, H. (1960/2010), Paradigms for a Metaphorology, Cornell UP
Clark, A. & Chalmers, D. (1998), “The Extended Mind,” Analysis 58
Friston, K. (2010), “The Free-Energy Principle,” Nature Reviews Neuroscience 11
Hayles, N.K. (1999), How We Became Posthuman, Chicago UP
Merleau-Ponty, M. (1945), Phenomenology of Perception
Nisbett, R. & Wilson, T. (1977), “Telling More Than We Can Know,” Psychological Review 84
Paolucci, C. (2024), Nati Cyborg, Luca Sossella
Parfit, D. (1984), Reasons and Persons, Oxford UP
Schwitzgebel, E. (2008), “The Unreliability of Naive Introspection,” Philosophical Review 117
Stiegler, B. (1994/1998), Technics and Time, 1, Stanford UP
Varela, F., Thompson, E., Rosch, E. (1991), The Embodied Mind, MIT Press
De Libera, A. (1996), La querelle des universaux, Seuil — on the medieval problem of universals.