This conversational essay revisits several themes developed in our article Understanding without Consciousness: Quantum Structures and the Autonomy of Meaning (published in Foundations of Science, 2026), to which we refer readers for a more in-depth treatment. Our central thesis is that understanding and consciousness need not necessarily coincide: to understand can mean being sensitive to the structure of meaning, its relations, its contextuality, and its latent possibilities, without this necessarily implying phenomenal access. From this perspective, we discuss the sense in which current language models may exhibit genuine forms of semantic intelligence and how the formalism of quantum cognition can help characterize certain structural properties of meaning.

Are we underestimating machines?

We would like to begin with a small defense of artificial intelligences. Not because we think they are poor creatures deserving our sympathy, of course, but because we are struck by the way we talk about them. We build them, train them, multiply them, continually change their architecture, use them to do an enormous amount of work, and then, when they display surprising capabilities, we immediately find a way to downplay them: “Yes, but they don’t really understand; they’re only machines, only statistics, only stochastic parrots.”

And the important word here is precisely “only.” Saying that an artificial intelligence uses statistical processes, neural networks, matrices, or mathematical rules may be perfectly correct. The problem arises when a description of its substrate is taken to have already explained the cognitive level. That would be like saying that a human being is only a collection of cells, or that a book is only paper and ink. It is not false, of course, but it is a description that fails to capture the level at which the phenomenon we want to understand appears.

And in our view there is also a form of anthropocentrism at work here. We humans are always very willing to recognize intelligence when we see it in our fellow humans, somewhat less willing when we see it in animals, and even less willing when it appears in an architecture we ourselves have built. As though the fact that we built the machine entitled us to think that everything it can do was already, somehow, trivially contained in the programmer’s instructions.

This is, after all, an old objection. As early as the nineteenth century, Ada Lovelace, while studying Charles Babbage’s mechanical computer, emphasized that machines could not originate ideas or autonomous thoughts because they simply carried out what we instructed them to do. A century later, however, Turing replied that a machine capable of learning might surprise us and produce results we had not anticipated. Today this discussion is no longer merely theoretical. We have systems that, starting from relatively simple training objectives, build internal representations far richer than anything their designers explicitly specified.

And this raises a question: when we say that a machine “doesn’t really understand,” what exactly are we saying? Do we have a positive criterion for distinguishing genuine understanding from merely apparent understanding? Or are we simply defending a boundary, saying: real understanding is the kind we achieve, and if a machine achieves it, then by definition it is simulated?

Our proposal is not to begin with a ready-made answer, but first to clarify what we mean by the term “understanding.” Above all, we should separate two questions that are often merged too quickly: the question of meaning and the question of consciousness.

The consciousness shortcut

A familiar example is Roger Penrose, who links three words: intelligence, understanding, and consciousness. Penrose says that to be intelligent one must understand, and to understand one must be conscious. If we accept this chain, then a machine, if it is not conscious, is excluded from genuine intelligence and understanding from the outset.

The problem is that consciousness then becomes a criterion of demarcation. But consciousness is precisely one of the things we understand least. If, in order to explain understanding, we invoke something even more mysterious, we have hardly made progress. More importantly, we have no operational criterion for verifying the presence of subjective experience from the outside. Each of us directly knows what it is like, for ourselves, to see red or feel pain. But we do not have the same kind of access to another person’s experience, and still less to that of an artificial system.

So we are not saying that consciousness does not exist or that it is unimportant. We are making a more modest claim: we cannot use it as a litmus test. We cannot say, “It understands because it is conscious, and it does not understand because it is not conscious.”

Turing saw this point clearly. If we demanded direct access to consciousness as proof of genuine thought, we would end up in solipsism: strictly speaking, we could no longer be certain that another subject possesses genuine understanding. In ordinary life, in fact, we attribute understanding to others primarily through what they do: their ability to respond to context, rephrase, correct themselves, grasp implications, metaphors, and figures of speech, and transfer knowledge from one situation to another.

John Searle then offers a more sophisticated objection with his Chinese Room. Let us briefly recall the thought experiment. Imagine a person who does not know Chinese, locked in a room with an enormous rulebook. Chinese symbols come in; the person applies the rules and sends other symbols out. From the outside, the answers may look correct, as though they had been generated by someone who understands Chinese, yet the person inside the room does not understand a word of Chinese. Therefore, Searle argues, syntax is not sufficient for semantics.

On this point we obviously agree with Searle: purely syntactic manipulation, carried out by applying predefined rules without an internal model of semantic structure, cannot be called understanding. But this is precisely where contemporary LLMs differ. The point is not that a string of symbols magically turns into meaning once it becomes long enough. The question is whether the system has managed to build an abstract, context-sensitive representation capable of organizing conceptual relations and latent possibilities. If it has, we are no longer in the rulebook scenario of the Chinese Room.

We can therefore say that the real opposition is not between human and machine. Nor is it between conscious and nonconscious. It is rather between “superficial symbol manipulation” and “structured semantic competence.” And this distinction already forces us to clarify three words that are often conflated: symbol, information, and meaning.

Symbols, information, and meaning

Let us make this distinction carefully. A symbol is something concrete and discriminable: a written letter, a sound, a gesture, an electrical signal, a physically realized bit. A must be distinguishable from B, and so on. Without differences there is no alphabet, and without an alphabet there is no message.

Then there is information. Claude Shannon is useful here. I can measure how much information a sequence of symbols contains without knowing what it means. I can talk about the probabilities of the symbols, their redundancy, the transmission capacity of a channel, and so on. None of this, however, tells me whether the sentence I am transmitting is a poem, an insult, a theorem, a recipe, or a meaningless string.

And then there is meaning. This is the most difficult level, because it is not simply another quantity that we can easily measure. It depends on context, potential relations, and emergent properties. The same word changes meaning in a different sentence. And the meaning of the whole is not simply the sum of the meanings of the parts.

We can use the idea of a “carrier.” A symbol can carry information; information can carry meaning. It is a little like physics, where a moving body carries momentum, and momentum carries energy. To say that one thing carries another does not mean reducing one to the other. Paper carries ink; the arrangement of the ink carries a text; the text carries a story. If I burn the sheet of paper, I burn the story’s physical support, but the story as a structure of meaning is not identical to the cellulose.

Consider, for example, a sad scene in a movie. It is a complex structure, full of meaning. But the director can use “the sad scene” as a single element within a larger narrative structure. Or consider a melody that, at one level, has musical meaning but, at another, can become the agreed-upon signal between two secret agents. The point is that what, at one level, is an entity of meaning can, at another level, function as a new symbol carrying information.

This also reminds us that when we say a neural network “only manipulates symbols,” it is important to ask at what level we are describing it. At a certain level, the brain too is made only of electrochemical events. But from this we cannot infer that our thought reduces to a description of those electrochemical events alone. This is the whole issue of emergence. The question to ask is: what organization emerges from the electrochemical level, and what does it enable?

We can therefore already formulate a first intuitive criterion: a system operates at the level of meaning if it does not respond only to the local form of the signs it uses, but is sensitive to a more abstract and emergent network of relations characterized by contextuality, potentiality, and nonclassical compositionality.

Structural meaning

Contextuality tells us that a word does not carry a meaning around with it like a rigid label. The meaning of a word such as the Italian word penna, for example, changes completely depending on the context. In a kitchen, penna can refer to a piece of penne pasta; for a writer, to a pen; for a bird, to a feather. Contexts, however, are often indeterministic, in the sense that different meanings can be actualized at different times in ways that are not predictable. This is the aspect of potentiality: there are latent meanings that can be made manifest, actualized, depending on the context. Finally, noncompositionality tells us that when we combine two concepts, the result may have properties that cannot simply be derived from the two concepts taken separately.

This is what we mean when we speak of structural meaning. “Structural” does not mean rigid like Lego bricks. We are talking about an abstract and extremely dynamic structure. As an image, we might think of an ecosystem: there are species, niches, dependencies, latent possibilities. Move one component and the functional value of many others changes. Or we can think of a space of social relations: the distance between two people is then not a distance in meters, but a distance determined by trust, affinity, and so on. One might be tempted to use the word “map,” but that risks suggesting something too static. It is better to think of a landscape of possibilities that contains not only what is currently present, but also what could become relevant if the context changed.

And here we begin to see why the vector spaces of neural networks matter for modeling meaning. In such a representation, a concept is not located at a single point as something isolated. It is defined by a set of relations, directions, proximities, and possibilities for transformation. We are not saying that having vectors and vector spaces is enough to capture a semantic space; we are saying that a relational geometry based on potentiality and contextuality is much closer to the kind of structure that meaning requires than a simple list of rules.

This is also why it makes much more sense to speak of understanding as a graded property than as an on/off switch. A system can grasp some aspects of a concept and miss others. It may be very powerful at grasping abstract linguistic structure and much less capable of handling the sensorimotor dimension. We ourselves understand differently depending on our accumulated experience, our cultural background, and so on.

So in our proposal, to “understand” means being sensitive enough to the structure of meaning to use it: to recognize relations, preserve coherence across multiple transformations, infer consequences, generate new combinations, and respond in contextually appropriate ways. Consciousness may accompany all of this, but it is not what defines this sensitivity to semantic structures.

From rules to a model

Let us try to explain how we arrived at all this. Take chess as an example, after first clarifying the historical stages. In 1997, IBM’s Deep Blue defeated Garry Kasparov. It was a very powerful system, with functions designed in part using human chess knowledge. It was already impressive, but it was still quite close to the classical idea of what is called an “expert system.” Many years later, however, AlphaZero was able to do something completely different. It was given only the rules of the game and one objective—to win—and it learned by playing against versions of itself. It was not handed, in the form of a manual, the strategies accumulated by chess culture. And yet positional evaluations, strategic concepts, and styles of play emerged that we recognize as intelligent, sometimes even original relative to human play.

What is important to emphasize is that the initial objective was extremely simple: “Win the game.” But to do this very well—that is, to win games even against the best world champions—the system had to build a model of the chess world. It had to learn that certain configurations have strategic value, that an apparently expendable piece can play a role, that the king is exposed, that a pawn structure contains future possibilities. None of these things are explicitly written in the rules of the game.

This is a perfect example of structural emergence. The rules of chess define a small universe. As soon as you define it, an enormous landscape of properties, relations, and strategies opens up. The same thing happens in mathematics. You define a structure and then discover that this definition has opened a conceptual territory far larger than the initial act itself. For example, when Évariste Galois defines the notion of a group, he is not simply inventing a word. He is making an entire domain of relations available. Mathematicians can then explore it, discover theorems, connections, and exceptions—in short, discover structure.

And here the old question immediately arises: did we create that structure, or did we discover it?

Probably both. A definition can generate a new conceptual landscape and, once created, some of its regions immediately become stable and well defined, independent of whether anyone is exploring them. But other regions remain potential, latent, and may require new creative acts to bring them into focus. For our argument, the main point is this: meaning can possess a coherence and autonomy of its own that do not coincide with the conscious act that originally generated it.

This also answers an objection often made to machines: “They create nothing; they only discover what was already predetermined.” But we cannot even draw such a clear line for human mathematics. If a machine finds a structure that no programmer encoded or anticipated, hypothesizing that it was already given, that it was predetermined, does not eliminate the fact that the system had to imagine and model it.

Above all, this reminds us not to confuse the simplicity of a learning rule with what results from putting that rule into practice.

Predicting the next word

This last point is crucial. With LLMs, one often hears the following caricature: the only thing they have learned to do is predict the next word (in fact, the next token); therefore, obviously, they cannot understand anything they write. But the reasoning can be reversed. Precisely because the initial task was to learn how to continue human texts across an enormous range of different contexts, the most effective way to do so was to learn to think like a human being—that is, to learn to understand human meaning.

If we write, “I left the ice in the sun, and after an hour...,” the next word depends on knowledge that is not purely grammatical. It depends on the fact that ice melts, on the context, on what is plausible. Multiply this by billions of situations and we can see that linguistic prediction pushes the system to compress an enormous amount of information. And such compression corresponds to a sensitivity to meaning.

In this sense, the statistical processes of modern LLMs are not alien to meaning. Quite the opposite. This is an important point to clarify. We often hear, “It understands nothing; it only does statistics.” But we humans also model statistical regularities. The question is: what kind of representation emerges from this statistical learning? If the system constructs latent spaces in which semantic relations are preserved at different levels of abstraction, then statistics, together with the mathematics of vector spaces, simply becomes the means through which a sensitivity to meaning takes shape.

Of course, this does not mean that an LLM understands everything, or that it understands the way a human being does. It may have gaps, hallucinate, lose coherence, or be fragile on certain tasks. But human understanding is also partial, graded, and distributed across very different competencies. The fact that a system makes a mistake does not show that it understands nothing, just as a student’s mistake does not show that the student has no understanding of the subject.

Another issue concerns novelty. Certainly, a random generator can produce a sentence that has never appeared before. Obviously, that is not enough. The criterion is whether the novelty we observe is semantically organized. If we ask a system to translate an unpublished poem while preserving an ambiguity, or to rewrite an argument from the opposite point of view without losing its structure, it is not enough to add new or surprising words. Meaning must remain invariant through complex transformations.

And this is one empirical way in which we can try to assess semantic competence: not merely by asking whether the system gives the right answer in certain cases, but by asking whether it preserves a conceptual structure, a domain of coherence, when the expressive form, context, perspective, language, level of abstraction, and so on change. The more robust and generative this coherence is, the less plausible it becomes to describe intelligent behavior as mere imitation of intelligence.

And this is where the word “simulation” becomes ambiguous. If I build an artificial hand that grasps, senses pressure, modulates force, compensates for errors, and performs all the relevant functions of a hand, at some point it becomes very strange to say that it “simulates grasping.” It actually grasps, even though its internal architecture is different.

A function can have multiple realizations. The platypus, to take a biological example, digests food despite lacking a stomach, and therefore does so with an apparatus very different from what we imagine as typical of a mammal. We cannot define a function by imposing a single architecture in advance. Likewise, an artificial semantic intelligence does not have to be a copy of the human brain in order to be a genuine form of semantic intelligence.

Mary’s room

At this point someone will say: all right, but experience is missing. A machine may know everything you like about red, but it does not know what it is like to see red. And here everyone thinks of Frank Jackson’s famous thought experiment, Mary’s room. Mary is a scientist who lives in a black-and-white environment and, by hypothesis, possesses all physical and functional knowledge about color vision. She knows what wavelengths do, how photoreceptors respond, how the brain processes color, how people classify and describe different colors, and red in particular. Then she leaves the room and sees red for the first time. Clearly something new happens: she acquires phenomenal access that she did not have before.

Our question is this: is that new experience foundational for understanding the meaning of the color red, or does it simply enrich an already existing structural understanding? We favor the second interpretation. Mary could already understand a great deal about red. When she sees it and experiences it, she adds a new mode of access to what she understands. She adds new information. And of course that access can further develop her semantic intelligence. An experience is not an irrelevant ornament. It can add associations, generate emotions, and promote new distinctions. But that is different from saying that no understanding existed before. This is precisely the distinction we are trying to draw between structural meaning and phenomenal access to meaning.

We find this distinction within ourselves as well. A great many cognitive processes occur without genuine conscious access. Think of a mathematician who works on a problem, then sets it aside, sleeps on it, and the next day an idea appears. Or of a writer who senses that a sentence does not work before being able to explain why. Cognitive processes can produce semantically valuable results without being present in the theater of consciousness.

Poincaré described precisely this kind of experience in mathematical creativity. Apparently, “the unconscious understands”: the equation “cognition equals consciousness” is therefore problematic.

We can also invoke the opposite example of meditation. When certain practices reduce the discursive and metacognitive flow, conscious experience does not necessarily disappear; indeed, for many practitioners it may become more evident precisely as conceptual thought decreases. This too suggests that consciousness and cognition are not synonyms.

Touching the world

Another strong objection concerns grounding, our anchoring in the world. It is often said that an LLM reads texts produced by human beings but does not see, touch, or have a body. Its meaning would therefore be derived, parasitic on ours.

This is a serious objection, but in our view not a decisive one, because human cognition itself has no direct access to reality. We access the world through receptors, neural transductions, perceptual categories, language, culture, and the testimony of others. When we read a book about Antarctica without ever having been there, our relationship to that part of the world is mediated. That does not make it meaningless.

We might say, provocatively, that human beings are part of an LLM’s sensory apparatus. Texts are traces produced by organisms that have interacted with the world. The model certainly does not have the same kind of access we do—not yet—but neither is it isolated from every interaction with reality.

Today, moreover, systems are becoming increasingly multimodal: images, audio, video, sensors, robotics. So this boundary is moving. But the philosophical point remains: we should not define understanding in such a way that only our biological mode of contact with the world can qualify as understanding. Language itself, moreover, is a vast technology of social grounding. Through language I can acquire information about events I have never perceived, periods in which I did not live, objects I cannot see directly, even mathematical structures that are not sensory objects. An essential part of human cognition is already mediated by other agents.

This brings us back to Mary’s room: direct experience and understanding are connected, but they do not coincide. Direct experience can enrich a model, but a model can exist and function through indirect access to reality as well.

Is every simulation fake?

The movie The Matrix seems to us a useful metaphor here. In The Matrix, you live in a simulated environment. Yet there you learn, suffer, love, make decisions, and build relationships. It would be too simple to say that “it is all fake.” It is false with respect to certain physical properties of the outside world, certainly, but it is real as an environment of interaction, with its own causes and effects. Part of the confusion comes from the fact that the word simulation covers very different cases. A weather simulation does not get anyone wet: it represents rain. A cardiac prosthesis, by contrast, does not merely “simulate” pumping blood; if it performs that function in the body, it actually pumps blood. So we have to ask: is the property we are discussing merely represented, or is it functionally instantiated?

In the case of semantics, if a system recognizes meanings, combines them, translates them, uses them to draw inferences, detects contradictions, changes its response as context changes, and produces new coherent structures, how can we continue to use the word “simulation”? To keep saying that it “only simulates,” we would need to provide an independent criterion of genuine understanding that the system fails to meet. If that criterion is simply “it must be conscious,” then we are back where we started. A useful and sensible criterion must be operational: it must be based on semantic tests and allow us to compare different degrees of competence.

This certainly does not erase the differences between a contemporary LLM and us humans. A current LLM does not possess continuous autobiographical memory, does not possess an individual body, and apparently does not possess persistent goals of its own. All of these things matter when we talk about agency or, more generally, personhood. But we are not claiming that an LLM is a person. We are claiming that an LLM is an entity endowed with semantic intelligence.

This is a fundamental distinction. Asking whether it understands is not the same as asking whether it is conscious, whether it has a self, whether it has desires, interests, moral responsibility, full autonomy, and so on. The debate about artificial intelligence often becomes confused because people try to answer all these questions with a single yes or a single no, conflating understanding and consciousness.

Quantum cognition

Perhaps it is time to introduce the topic of quantum theory. When we say “quantum” in relation to cognition, many people immediately assume that we are talking about the brain, or a computer, as physical quantum entities in states of superposition. But that is not the perspective we adopt: ours is the perspective known as quantum cognition.

Quantum cognition is a research program that uses the mathematical formalism of quantum theory to model cognitive phenomena. It emerged because human decisions, concept combinations, order effects, and contextuality often violate classical probabilistic models, yet are described very effectively by quantum-like structures. Here, the term “quantum structure” describes a very particular form of organization that can be present in very different kinds of systems, not a microscopic mechanism located, for example, in the brain.

Before being subjected to a questioning context, a concept can be represented as a superposition of possibilities. The context then functions like a quantum measurement: it does not merely read out an already predetermined property, but helps actualize one among many possibilities, without any of them being given a priori. And when we combine concepts, the new composite conceptual entity can give rise to quantum-like correlations that violate the famous Bell inequalities and cannot be decomposed in classical terms in the way correlations between separate systems can.

All of this connects perfectly with the properties we have discussed in relation to meaning: contextuality, potentiality, and noncompositionality.

And this is the deeper reason why quantum formalism is interesting for modeling meaning. Not because meaning is made of microscopic particles, but because quantum formalism was built precisely to describe potential properties, transitions from the potential to the actual, the corresponding probabilities, the dependence of properties on the measurement context, correlations that cannot be reduced to common causes in the past, and so on. Meaning seems to have all of these characteristics.

Here again, it is very important not to confuse information and meaning. In physics we speak of “quantum information,” but the expression has a precise technical meaning. In quantum cognition, by contrast, the formalism is used to model not information, but semantic structures. If we call everything information, we risk losing the important distinction we made at the beginning.

Our central argument—that understanding can exist in the absence of consciousness—obviously does not depend on the successes of quantum cognition. If a better formalism for describing semantics were discovered tomorrow, our thesis would remain valid. The successes of quantum cognition simply provide an additional confirmation by showing that meaning possesses structural regularities robust enough to be modeled mathematically through quantum formalism.

A gas of words

Let us consider an important example of quantum-like structural regularities: the statistics of words in a natural-language text. The idea is to describe the text as a “gas of words,” understood as entities that carry meaning. We then rank the different words by frequency of occurrence. In this model, the most frequently used word occupies the lowest energy level; the second most frequent occupies the next level, and so on. The different occurrences of the same word are treated as indistinguishable entities forming this strange gas of words.

At this point we can ask: what statistics describe the distribution of these frequencies? Classical statistics, such as Maxwell-Boltzmann, or quantum statistics? Aerts and Beltran discovered a few years ago that human texts are described remarkably well by the Bose-Einstein distribution, and therefore by bosonic statistics, in a kind of generalization of the famous Zipf law.

What is especially interesting for our discussion is that the same “bosonic signature” also appears in texts generated by LLMs. And not because they literally copy human texts, since these are generated texts. The relative frequencies of the words used continue to organize themselves according to a structure of the same type.

We can therefore observe a convergence between human cognition and artificial cognition. Two systems with completely different histories and substrates end up producing linguistic entities that display the same statistical-semantic organization. This does not prove that they think exactly the way we humans do in every detail, but it is an important indication that both operate under common structural constraints characteristic of this strange substance we call meaning.

It is somewhat as if two species that evolved separately developed similar solutions because they face the same kind of problem. Here the problem is how to move effectively through the vast landscape of language and meaning.

And the solutions in question bear a quantum signature. The linguistic traces produced by LLMs display quantum-like regularities equivalent to the traces we humans leave. Of course, let us repeat once again to avoid any misunderstanding: we are not saying that words in a text are physical bosons. We are saying that the mathematical organization of their relative frequencies is described by a bosonic structure, similar to a Bose-Einstein condensate.

We can use the example of an ocean wave and an electromagnetic wave: both are described by the same wave mathematics without therefore being the same thing. They share a wave-like nature, but they are very different entities.

Is consciousness quantum?

But now it is worth discussing a risk. If we see quantum structures in cognition and meaning, we might be tempted to say: there it is—quantumness and consciousness are the same thing. Or: every quantum state possesses an inner, experiential aspect.

To be clear, this is a philosophical possibility, advocated, for example, by D’Ariano and Faggin. But it is absolutely not a consequence of the results of quantum cognition. Quantum models show that we can describe cognitive and semantic structures very well, but they do not show in any way that such structures are associated with phenomenal experience. Of course, if we conflate meaning and consciousness, such an association might seem natural. But it is precisely this ad hoc association that we contest. One may be interested in panpsychism or in proposals such as those of D’Ariano and Faggin, but it is not enough to say: there is quantum formalism, therefore there is consciousness.

The formalism may be a signature of the presence of cognitive-conceptual structures without necessarily being a signature of conscious experience. Conscious experience, when it is present, may inhabit meaning, but a cognitive entity’s sensitivity to the structure of meaning is not equivalent to its phenomenal access to it.

One can be semantically intelligent without being conscious.

This caution, this demarcation, actually protects consciousness as well. If we use consciousness to explain every aspect of meaning, we risk diluting it. It is far more interesting to leave the phenomenal mystery open and separately study what we can characterize at the cognitive and semantic levels.

And the Chinese Room?

Let us return now to the Chinese Room and see whether, after this journey, we can look at it differently. In the original scenario there is a rulebook that formally associates inputs and outputs. If the system truly constructs no abstract model of the language, then we agree: it does not understand Chinese.

The question is: can a rulebook capable of sustaining an open-ended, contextual, creative conversation—translating metaphors, grasping irony, learning new conventions, generalizing, and correcting itself—still be thought of as a simple list of rules? At some point the “rulebook” would have to incorporate a structure so rich that, functionally, it would be a model of the language and the world.

It is almost as if the thought experiment hides all the hard work inside the rulebook while pretending that the rulebook is simple. It says: imagine a book of rules that produces responses indistinguishable from those of a competent person. But how is that book organized? If it really could do this in any context, it would have to contain precisely the semantic structure that the argument aims to exclude.

This does not refute Searle on every level, but it certainly weakens the use of the Chinese Room as a demonstration that computation can never access semantics. It depends on what we call computation and on what abstract representations emerge in the system.

This forces us not to fall into the trap of focusing on the wrong level. Saying that an LLM contains matrix multiplications is like saying that the brain contains ionic exchanges. It is true, but cognition has to do with the organization realized by those dynamics. It exists at another level.

The substrate level is necessary but not sufficient to explain the semantic level. Meaning resides in the organized relations that the substrate makes possible, not in the individual transistor, the individual neuron, or the individual isolated word.

Perhaps we can therefore say that the Chinese Room remains an excellent warning against the naive idea that syntax and semantics are the same thing. But it is not a good reason to deny a priori that an artificial system can construct a genuine semantic dimension.

That is exactly the position we defend.

Understanding without experience

If we had to condense all of this into one sentence, we would say: perhaps we have made the mistake of calling “understanding” only that portion of meaning to which a subject has conscious access. But the domain of meaning is much larger than the domain of conscious experience.

More precisely, we would formulate it this way: what is phenomenal is not meaning as such, but the way a subject accesses that meaning. Meaning is structural, relational, contextual, potential. Consciousness concerns the experience that may accompany or inhabit that organization.

From this perspective, LLMs become an extraordinary philosophical and scientific experiment. Not because they are already artificial persons, and not because we have shown that they are conscious (we obviously make no such claim), but because they display a highly advanced form of semantic competence in a system for which we have no clear reason to posit a phenomenal experience comparable to our own.

They force us to break an old association that, in our view, is mistaken: the idea that intelligence and understanding coincide with the phenomenon of consciousness. We can have semantic intelligence without being conscious. And we can study semantic intelligence by exploring its capacity to generate coherence, operate on latent possibilities, and produce new configurations of meaning.

This obviously does not diminish the human being. We would say the opposite. If we recognize that an enormous part of cognition is structural, we can finally ask more precisely what consciousness adds, studying it as a specific phenomenon rather than as a synonym for every mental phenomenon.

Nor does it diminish machines. The word “machine” often makes us think of a mechanical clock, of rigid and deterministic gears. But contemporary physics and systems theory have taught us that material structures can support extraordinarily rich dynamics. Saying that something is a machine does not yet tell us what can emerge from its organization.

Perhaps the question we should ask is no longer whether machines understand or not, but what forms of understanding exist. Are there dimensions of human understanding that artificial systems do not possess, or never will? Which dimensions, by contrast, might machines develop better than we do? In short, understanding not as a simple binary property, but as a spectrum of possibilities that can be explored by both biological and artificial systems. And consciousness not as a mandatory ingredient of an intelligent system.

And at this point we can also make peace with the poor stochastic parrots. They may be stochastic, they are certainly machines, and they are undoubtedly different from us. But none of these facts demonstrates that they do not understand. If we want to deny them understanding, we must do so by showing clearly where and how their sensitivity to meaning fails.