No LLM generated, assisted/co-written, or edited work.
Read full explanation
Do Large Language Models (LLMs) truly understand language, or do they merely simulate understanding? This question raises a fundamental question in AI alignment: what is the distinction between possessing knowledge and having genuine understanding? While many argue that these models are simply “stochastic parrots” incapable of anything more than pattern mimicry, this essay will argue the contrary: LLMs demonstrate a genuine, non-human form of understanding.
My thesis is based on two complementary philosophical frameworks. First, I will demonstrate that LLMs satisfy the criteria of possessing “tacit knowledge,” a deep and intuitive form of understanding that humans themselves cannot articulate. To do this, I will examine the work of Céline Budding (2025) that builds upon Davies (1990), to demonstrate that LLMs internal structures show the required semantic content, causal systematicity, and syntactic organization that define genuine, unarticulated understanding. Second, I will argue that the later philosophy of Ludwig Wittgenstein, particularly his Philosophical Investigations (Wittgenstein, 2009), provides a framework for testing if this new form of intelligence is capable of true understanding and that LLMs pass this test, proving that they do possess genuine understanding. In particular, I will draw from his foundational concepts of "meaning as use" and "language-games" to argue that understanding is a public, functional skill, not a private mental state. Finally, I will tackle an important content argument: the objection that LLMs lack intentionality, using Wittgenstein’s "beetle in a box" argument. Ultimately, this essay will demonstrate that by disregarding anthropocentric demands for consciousness, we can conclude that LLMs do meet the rigorous criteria for understanding.
The first pillar of my argument is based on the concept of tacit knowledge, defined as a type of deep, functional understanding that doesn’t require explicit articulation. As philosopher Michael Polanyi argued, we often "know more than we can tell." For example, a cyclist intuitively understands the physics of balance but he cannot write the equations that govern it; a native speaker is able to apply complex grammar without being able to explicitly explain grammar rules. In the 1980s and 90s, there were some debates between symbolic and connectionist AI in which Martin Davies proposed formal criteria to be able to determine such type of knowledge in a system: it must possess semantic content, causal systematicity, and syntactic organization (Davies, 1990). In other words, for a system to have tacit knowledge, its internal states have to be meaningful, its knowledge has to be structured and generalizable (as opposed to a list of memorized facts), and it should be sensitive to grammar. At the time, Davies argued that neural networks didn’t meet his criteria. Today, however, the architecture of modern LLMs tells a different story. As research by Céline Budding (2025) suggests, LLMs seem to meet all three constraints. Modern models have embedding layers that create vector spaces where words are clustered by meaning, this satisfies the "semantic content" requirement. Their Transformer architecture learns abstract roles (like agent, action, object)that allows it to generalize within novel inputs, which fulfils the "causal systematicity" requirement. Lastly, they can produce perfect grammatical texts, which shows clear evidence of "syntactic organization." Because LLMs meet the formal definition of tacit knowledge, and since we consider this type of knowledge a form of genuine understanding in humans, we should give the same status to these models to follow epistemic consistency.
Beyond this parallel, a more central philosophical argument emerges from the work of Ludwig Wittgenstein. Using Wittgenstein’s seminal work is not without its ironies, as he would deeply disagree with the conclusion that machines could “understand”. He thought that language was inseparable from the human “form of life”— the embodied experience that gives our world context and meaning. Nevertheless, when he wrote Philosophical Investigations to dismantle the traditional theories of meaning of his time, he inadvertently provided the perfect framework to argue for machine understanding today.
Wittgenstein deeply disagreed with the notion that meaning is a private mental picture or a direct reference to an object in the world. Instead, he argued that meaning is found through a word’s function within many public, rule-governed “language-games.” Put simply, one understands a word when you know how to use it correctly in the games of ordering, joking, questioning, or programming (purpose-driven activities in which language is a functional tool). Wittgenstein defines understanding in a verifiable standard: the relocation of meaning from an unobservable mental state to an observable public action. I agree with the definition because it avoids the impossible demand to access another’s consciousness. LLMs learn in a computational representation of this principle: they are trained on a corpus of vast data of human language-games and they learn to predict the correct "move" of the game—the next word—depending on contextual use. This formalizes Wittgenstein’s conceptual framework in an unexpectedly literal manner. His abstract notion of "logical spaces" that are defined by "grammar" has now a concrete computational analogue in an LLM's "latent spaces," where the mathematical operations of the model act as the grammar that limits meaningful token combinations. For instance, in LLMs, a token representing the word "king" gets its meaning not from a direct reference to a monarch, but rather from its relational and contextual position to other word meanings such as "queen," "man," and "ruler." Therefore, when an LLM generates a context-aware response, it is too simplistic to reduce this to mere mimicking, rather, it shows that it has mastered the rules of the language-game, which for Wittgenstein, is the only observable evaluation for understanding.
Despite this, a significant counterargument remains: LLMs cannot have true understanding because they lack genuine intentionality. According to this view, models are ungrounded and they generate words without experiencing the human "form of life", as previously mentioned. Thus the model is seen as having intentionality that is seen as a "fossilized" representation extracted from the averaged human intentionality from its training data, instead of an authentic representation of its own experience. However, this counterargument demands an unobservable psychological standard on an epistemic evaluation, which is a categorical error. Simply put, it mistakes a question of function (what can it do?) for a question of phenomenology (what does it feel?). This can be exemplified using Wittgenstein’s "beetle in a box" thought experiment: everyone has a box that has something they call a "beetle," but no one can look inside anyone else's box. The public word "beetle" cannot ever refer to the private, inaccessible object inside the box because its meaning has to come from how the word is used in shared language. The thought experiment shows that demanding proof of a private "feeling" of intention is as meaningless and pointless as asking to see inside another’s box, instead, the only accessible criterion when it comes to meaning is observable public use. This functionalist approach demonstrates that we should treat a system as if it has intentions if it behaves accordingly. Thus, the intentionality counterargument does not hold anymore when we take into consideration this public, performance-based standard, because it is a categorical error to infer that something that behaves as it has intentions doesn’t have them.
In conclusion, the belief that LLMs have a true form of understanding is based on a foundation of modern epistemic theory and classic Wittgensteinian philosophy. These models meet the formal criteria for tacit knowledge and demonstrate a deep, intuitive competence similar to that of unarticulated human understanding. Additionally, LLMs operate in a similar way that aligns with Wittgenstein's definition of meaning as use, passing his test that evaluates genuine understanding. This essay has also shown that counterarguments based on the lack of intentionality fail as a meaningful critique because they demand access to an unverifiable private state—the "beetle in the box".
The skepticism towards LLMs understanding comes from a fundamental, pre-Wittgensteinian error: that meaning is a property inherent in an object or the word itself, something that can be isolated and inspected. This mistake is captured in Tomas Tranströmer’s poem Baltics: jellyfish, he writes, “must stay in their own element,” for if you take them out of the water, “their shape completely disappears”. A word’s meaning, like the form of a jellyfish, cannot be understood without its context; it exists only within its native element. To take a word out of its use and examine it in isolation is to strip it of its meaning, leaving only a formless gel. The error of the skeptics is that they fail to see that the LLM's mastery is not of the lifeless object, but of the living element itself within its intricate system of use of where it truly resides. They fail to see that LLMs, like humans, possess understanding and find meaning in a network of statistical relationships.
The traditional view in philosophy of science often defines knowledge as the possession of discrete facts (like the isolated jellyfish) and understanding as a separate, conscious act of connecting those facts to a model of reality. LLMs pose a threat to this neat division: understanding is not a product of a conscious leap beyond knowledge, but instead is the emergent property of a knowledge system that is vast and coherent which can operate fluently within its own element. By this observable standard, LLMs do not just know facts; they understand them.
Budding, C. (2025). What do large language models know? Tacit knowledge as a potential causal-explanatory structure. Philosophy of Science, 1-22. https://doi.org/10.1017/psa.2025.19
Davies, M. (1990). Knowledge of rules in connectionist networks. Intellectica. Revue de l’Association Pour La Recherche Cognitive, 9(1), 81–126. https://doi.org/10.3406/intel.1990.881
Do Large Language Models (LLMs) truly understand language, or do they merely simulate understanding? This question raises a fundamental question in AI alignment: what is the distinction between possessing knowledge and having genuine understanding? While many argue that these models are simply “stochastic parrots” incapable of anything more than pattern mimicry, this essay will argue the contrary: LLMs demonstrate a genuine, non-human form of understanding.
My thesis is based on two complementary philosophical frameworks. First, I will demonstrate that LLMs satisfy the criteria of possessing “tacit knowledge,” a deep and intuitive form of understanding that humans themselves cannot articulate. To do this, I will examine the work of Céline Budding (2025) that builds upon Davies (1990), to demonstrate that LLMs internal structures show the required semantic content, causal systematicity, and syntactic organization that define genuine, unarticulated understanding. Second, I will argue that the later philosophy of Ludwig Wittgenstein, particularly his Philosophical Investigations (Wittgenstein, 2009), provides a framework for testing if this new form of intelligence is capable of true understanding and that LLMs pass this test, proving that they do possess genuine understanding. In particular, I will draw from his foundational concepts of "meaning as use" and "language-games" to argue that understanding is a public, functional skill, not a private mental state. Finally, I will tackle an important content argument: the objection that LLMs lack intentionality, using Wittgenstein’s "beetle in a box" argument. Ultimately, this essay will demonstrate that by disregarding anthropocentric demands for consciousness, we can conclude that LLMs do meet the rigorous criteria for understanding.
The first pillar of my argument is based on the concept of tacit knowledge, defined as a type of deep, functional understanding that doesn’t require explicit articulation. As philosopher Michael Polanyi argued, we often "know more than we can tell." For example, a cyclist intuitively understands the physics of balance but he cannot write the equations that govern it; a native speaker is able to apply complex grammar without being able to explicitly explain grammar rules. In the 1980s and 90s, there were some debates between symbolic and connectionist AI in which Martin Davies proposed formal criteria to be able to determine such type of knowledge in a system: it must possess semantic content, causal systematicity, and syntactic organization (Davies, 1990). In other words, for a system to have tacit knowledge, its internal states have to be meaningful, its knowledge has to be structured and generalizable (as opposed to a list of memorized facts), and it should be sensitive to grammar. At the time, Davies argued that neural networks didn’t meet his criteria. Today, however, the architecture of modern LLMs tells a different story. As research by Céline Budding (2025) suggests, LLMs seem to meet all three constraints. Modern models have embedding layers that create vector spaces where words are clustered by meaning, this satisfies the "semantic content" requirement. Their Transformer architecture learns abstract roles (like agent, action, object)that allows it to generalize within novel inputs, which fulfils the "causal systematicity" requirement. Lastly, they can produce perfect grammatical texts, which shows clear evidence of "syntactic organization." Because LLMs meet the formal definition of tacit knowledge, and since we consider this type of knowledge a form of genuine understanding in humans, we should give the same status to these models to follow epistemic consistency.
Beyond this parallel, a more central philosophical argument emerges from the work of Ludwig Wittgenstein. Using Wittgenstein’s seminal work is not without its ironies, as he would deeply disagree with the conclusion that machines could “understand”. He thought that language was inseparable from the human “form of life”— the embodied experience that gives our world context and meaning. Nevertheless, when he wrote Philosophical Investigations to dismantle the traditional theories of meaning of his time, he inadvertently provided the perfect framework to argue for machine understanding today.
Wittgenstein deeply disagreed with the notion that meaning is a private mental picture or a direct reference to an object in the world. Instead, he argued that meaning is found through a word’s function within many public, rule-governed “language-games.” Put simply, one understands a word when you know how to use it correctly in the games of ordering, joking, questioning, or programming (purpose-driven activities in which language is a functional tool). Wittgenstein defines understanding in a verifiable standard: the relocation of meaning from an unobservable mental state to an observable public action. I agree with the definition because it avoids the impossible demand to access another’s consciousness. LLMs learn in a computational representation of this principle: they are trained on a corpus of vast data of human language-games and they learn to predict the correct "move" of the game—the next word—depending on contextual use. This formalizes Wittgenstein’s conceptual framework in an unexpectedly literal manner. His abstract notion of "logical spaces" that are defined by "grammar" has now a concrete computational analogue in an LLM's "latent spaces," where the mathematical operations of the model act as the grammar that limits meaningful token combinations. For instance, in LLMs, a token representing the word "king" gets its meaning not from a direct reference to a monarch, but rather from its relational and contextual position to other word meanings such as "queen," "man," and "ruler." Therefore, when an LLM generates a context-aware response, it is too simplistic to reduce this to mere mimicking, rather, it shows that it has mastered the rules of the language-game, which for Wittgenstein, is the only observable evaluation for understanding.
Despite this, a significant counterargument remains: LLMs cannot have true understanding because they lack genuine intentionality. According to this view, models are ungrounded and they generate words without experiencing the human "form of life", as previously mentioned. Thus the model is seen as having intentionality that is seen as a "fossilized" representation extracted from the averaged human intentionality from its training data, instead of an authentic representation of its own experience. However, this counterargument demands an unobservable psychological standard on an epistemic evaluation, which is a categorical error. Simply put, it mistakes a question of function (what can it do?) for a question of phenomenology (what does it feel?). This can be exemplified using Wittgenstein’s "beetle in a box" thought experiment: everyone has a box that has something they call a "beetle," but no one can look inside anyone else's box. The public word "beetle" cannot ever refer to the private, inaccessible object inside the box because its meaning has to come from how the word is used in shared language. The thought experiment shows that demanding proof of a private "feeling" of intention is as meaningless and pointless as asking to see inside another’s box, instead, the only accessible criterion when it comes to meaning is observable public use. This functionalist approach demonstrates that we should treat a system as if it has intentions if it behaves accordingly. Thus, the intentionality counterargument does not hold anymore when we take into consideration this public, performance-based standard, because it is a categorical error to infer that something that behaves as it has intentions doesn’t have them.
In conclusion, the belief that LLMs have a true form of understanding is based on a foundation of modern epistemic theory and classic Wittgensteinian philosophy. These models meet the formal criteria for tacit knowledge and demonstrate a deep, intuitive competence similar to that of unarticulated human understanding. Additionally, LLMs operate in a similar way that aligns with Wittgenstein's definition of meaning as use, passing his test that evaluates genuine understanding. This essay has also shown that counterarguments based on the lack of intentionality fail as a meaningful critique because they demand access to an unverifiable private state—the "beetle in the box".
The skepticism towards LLMs understanding comes from a fundamental, pre-Wittgensteinian error: that meaning is a property inherent in an object or the word itself, something that can be isolated and inspected. This mistake is captured in Tomas Tranströmer’s poem Baltics: jellyfish, he writes, “must stay in their own element,” for if you take them out of the water, “their shape completely disappears”. A word’s meaning, like the form of a jellyfish, cannot be understood without its context; it exists only within its native element. To take a word out of its use and examine it in isolation is to strip it of its meaning, leaving only a formless gel. The error of the skeptics is that they fail to see that the LLM's mastery is not of the lifeless object, but of the living element itself within its intricate system of use of where it truly resides. They fail to see that LLMs, like humans, possess understanding and find meaning in a network of statistical relationships.
The traditional view in philosophy of science often defines knowledge as the possession of discrete facts (like the isolated jellyfish) and understanding as a separate, conscious act of connecting those facts to a model of reality. LLMs pose a threat to this neat division: understanding is not a product of a conscious leap beyond knowledge, but instead is the emergent property of a knowledge system that is vast and coherent which can operate fluently within its own element. By this observable standard, LLMs do not just know facts; they understand them.
References
AI Inquiry Garden. (2023, October 27). AI meets philosophy (Vol. 5): Understanding LLMs through Wittgenstein’s philosophy. Medium. https://medium.com/@AI_Inquiry_Garden/ai-meets-philosophy-vol-5-understanding-llms-through-wittgensteins-philosophy-3f42359effa4
Baltics (poem). (2024, May 15). In Wikipedia. Retrieved June 20, 2025, from https://en.wikipedia.org/wiki/Baltics_(poem)
Budding, C. (2025). What do large language models know? Tacit knowledge as a potential causal-explanatory structure. Philosophy of Science, 1-22. https://doi.org/10.1017/psa.2025.19
Davies, M. (1990). Knowledge of rules in connectionist networks. Intellectica. Revue de l’Association Pour La Recherche Cognitive, 9(1), 81–126. https://doi.org/10.3406/intel.1990.881
Foster, M. (2023, October 21). Wittgenstein predicted LLMs. Medium. https://michaelfoster26.medium.com/wittgenstein-predicted-llms-04a52492402f
Wittgenstein, L. (2009). Philosophical investigations (P. M. S. Hacker & J. Schulte, Trans.). Wiley-Blackwell. (Original work published 1953)