In 2022, when Blake Lemoine shared his experience with LaMDA 2, he took the AI’s responses seriously and claimed that artificial intelligence (AI) consciousness is real. Imagine the movie Her, in which a man falls in love with an operating system he converses with. With the growing sophistication of LLMs (large language models) and their conversational patterns, the temptation to attribute human characteristics to them no longer seems dystopian. However, while the question of what kind of entity AI is currently occupies the central stage of scientific and public debate, Lemoine’s claim needs to be carefully substantiated before it can be considered true. As Emily M. Bender and Alexander Koller point out in their position paper, an assessment of what AI is must draw an adequate distinction between language and meaning. If this distinction is successful, as demonstrated by some thought experiments in the paper, it will become evident that language is nothing more than a formal structure, whereas meaning implies an understanding of the relationship between objects and context.
The belief that use and understanding are closely connected terms in the context of language has become popular among analytic philosophers of language, particularly since Wittgenstein, for whom understanding is rule-following behavior within a particular socio-cultural context; and although he mentioned psychical accompaniments, needless to say, he did not attach much importance to them. The growing complexity of the problems raised by artificial intelligence leads us to think that perhaps it is time to look back and break this hegemony of Wittgenstein’s view. In fact, an explanation of the nature of language in which knowledge and use are not necessarily linked was put forward in the past by Augustine of Hippo (354–430 CE). He argued that it is possible to use language in a justified manner without the user acquiring any knowledge from it.
Augustine took this thesis even further, developing a theory of language that led him to conclude that when we associate linguistic forms or signs, even when this association is derived from the context, this does not generate any teaching or learning. In his dialogue “On the Teacher,” he discussed the problem of the origin of learning and argued that human knowledge is fundamentally linked to the inner mind. For Augustine, signs, whether natural or linguistic, are external to the mind and are primarily ostensive in nature. Their function is to point towards another reality distinct from themselves. Although they possess significant potential, signs mean nothing by themselves. Therefore, even when used in communication processes, linguistic signs cannot convey knowledge. Now, this seems to be the characteristic of LLMs.
Words Without Knowledge in LLMs
An LLM works by receiving a prompt, which is a text input, and generating a response. It predicts text based on patterns it has learned by reading a massive amount of information during the programming process. During this training, it absorbs statistical relationships between words, axioms, and concepts. Contemporary researchers such as Luciano Floridi, Murray Shanahan, Kyle McDonell, and Laria Reynolds, and Rasmus Gahrn-Andersen are convinced that the functional architecture of LLMs, based on symbol formatting and statistical patterns, lacks sensory, social, and experiential groundings. The architecture does not share the life conditions that make expressions meaningful and do not form the kinds of causal/common-sense models that humans rely on; consequently, LLMs lack genuine understanding, since language use and knowledge are interconnected concepts, and language use based on words or signs requires an embodied and enactive foundation.
My co-author and I, in light of what these scholars have already argued, certainly assert that the LLM does not understand. However, unlike the proponents of the previous view, what we intend to argue is that the use of words and the understanding/knowledge of them are not the same activity—they are distinct operations. LLMs can be understood merely as operators or sophisticated users of signs, capable of producing an almost infinite number of signs with extraordinary efficiency. However, if we apply the basic principles of Augustine’s theory of language, we must conclude that this does not make them capable of understanding the meaning of what they are producing. Understanding and knowing are activities defined by components entirely different from signs and their combinations of usage. Given the sophistication of the technical infrastructure of LLMs, such combinations may appear to be knowledge. But in light of Augustine’s distinction between language and meaning, we must, in fact, uphold that LLMs, because they lack the internal mental state of knowledge, are not producers of meaning. According to Augustine, knowledge originates in the human mind, does not depend on signs, and is always accompanied by self-awareness.
Why Words Fail: Augustine and the Use of Language
Augustine asserted that the understanding of human language is not a mechanism resulting from some kind of order generated by signs, whether natural or conventional. Thus, when we speak of meaning, we always refer to an action of an inner faculty of the subject. This faculty, in turn, operates in a complex manner, interacting both with sensory representations and with the pure intelligible, but always from the inner self. Augustine demonstrated that language is a human act that consists in the construction of meaning from knowledge. Knowledge, in turn, is not generated through interaction with signs but rather through an action of truth, which is a principle that resides in and acts within the human mind. This action of truth in the human mind, which Augustine called illumination in “On the Teacher,” is the sole origin and cause of learning and understanding; words and all kinds of signs have no power in themselves to generate either knowledge or meaning. The power of signs is merely referential, that is, it is a power to point to or draw the mind’s attention to other objects.
What seems relevant to us in Augustine’s theory of language and in his description of the learning process is a twofold process: the combination of signs, especially conventional ones, and the production of knowledge. In the first case, we are dealing with a process that is not associated with any production of meaning. In the second, meaning is generated by the cognitive agent and, at a subsequent and secondary moment, is expressed in external signs.
Comparing LLMs with Augustine’s model is unusual, since his theory does not appear to be complementary. However, we can create a conceptual framework for LLMs based on Augustine’s theory of signs, which can support the idea that there is a mechanical, non-human use of signs that is not accompanied by the decoding of meaning and the production of knowledge. If we compare this perspective with the LLM, we can easily identify the commonalities between the two models of language use.
First, Augustine argues that the function of language consists in various signs that direct our attention but which cannot convey knowledge. Second, true understanding originates in the subject’s internal cognitive faculties. In fact, I cannot know that a sign is a sign unless I know what it means. But if meaning is a product of an inner faculty of the mind, the sign by itself does not generate meaning and is not a source of learning. Third, since nothing can be taught or learned through signs, if we define language as the use of signs, we are faced with an epistemologically empty product whose utility is merely instrumental.
Now, if LLMs function by receiving a prompt and predicting the most likely token from the vast corpus of data and instructions stored within them, yet without understanding the meaning of the signs, we must acknowledge that when we classify LLMs as language models, we are using this term to denote only the instrumental value of the tokens that LLMs possess. In fact, these entities we call LLMs use tokens to predict other tokens, but they lack understanding and are incapable of producing meaning, given that knowledge does not arise from mere rule-following behavior but from the inner workings of the mind, which they do not possess. If we consider Augustine’s distinction between the instrumental function of signs and the production of meaning by the mind, the latter being the very essence of language, we can say that LLMs are a technological simulation of that function of language use, but there is no production of any knowledge and understanding.
Augustine explains the activity of knowledge through his famous theory of illumination. Of course, the theory of illumination takes the important point from his discussion, but we can discuss his theory without referring to illumination, which is, for him, Christ as the Teacher operating within us. His theory provides a deeper insight into the knowledge, learning, and understanding that traditional analytic philosophy of language has ignored. The philosophers mentioned above, who argued for the impossibility of understanding and knowledge in LLMs, raise the hypothesis that mere rule-following, the manipulation of symbols, and statistical analysis cannot explain true comprehension. It is clear that they were influenced by the later Wittgenstein, and we do not deny that the use of language requires this lived experience, but understanding primarily arises from within. No one can ever know what a sign means merely from the sign itself, unless they were already aware of it through their inner reason. And this is an inner mental process that the Wittgensteinian tradition overlooked.
From the Augustinian perspective, it does not matter whether they have internal representations of rules or how they are following them, or whether they are aligned with the socio-cultural context or not; the fact is, LLMs possess no internal mechanism to understand them. Understanding occurs only through the intervention of reason because knowing something is only possible when we know why we know this or that. Even when LLMs follow the rules, they simply do not know what they are doing. As Peter King observes, for Augustine, obeying the rule is like being a parrot, which constitutes an argument similar to that of John Searle’s Chinese room or the stochastic parrots of Emily M. Bender, Timmit Gebru, Angelina McMillan-Major and Shmargaret Shmitchell. Knowing or understanding is an internal episode; now, this could be illustrated in multiple ways, through subjectivity or consciousness, or as something else, even if we do not subscribe to a theory of illumination.
Rituparna Roy
Rituparna Roy is a visiting researcher at the Institute of Philosophy of the University of Porto (Portugal) and is working with Paula Oliveira e Silva, principal investigator, on the project “Language and Knowledge: From Augustine to Contemporary Phenomenology.” Her expertise is in philosophy of mind and consciousness and proto-phenomenology. She also works on philosophy of language and Buddhist psychology. She did her PhD in the unity of consciousness at Jadavpur University, India. She also worked as a guest researcher at the Center for Mind and Cognition, Ruhr-University Bochum, with Tobias Schlicht, principal investigator. She presented several papers in both Europe and Asia in international conferences on the philosophy of mind, bringing together analytic and historical perspectives. Her most recent publication is “I Feel Like I Am Unable to Talk-Is Phenomenal Concept Effeable?” (2026).
Paula Oliveira e Silva
Paula Oliveira e Silva is Associate Professor at the Philosophy Department of the University of Porto (Portugal). She works in philosophy, with expertise in Augustinian Studies and an emphasis on metaphysics and theories of knowledge in the Western tradition. She was principal investigator of the research project “‘Animal Rationale Mortale’: The Mind-Body Problem and the Emotions in the Commentaries on Aristotle's De anima Produced in the 16th Century Portuguese Universities,” and she collaborated as a member of several research projects such as “From Data to Wisdom: Philosophizing Data Visualizations in the Middle Ages and Early Modernity.” She publishes on a regular basis on Augustine’s metaphysical theories of knowledge and freedom, her most recent book being As partes e o todo: ensaios sobre a filosofia de Agostinho de Hipona (2025).
