Can Large Language Models Understand Meaning? Nietzsche, Bilingual Thought, and Linguistic Form

Introduction: Can Language Models Have Meaning?

A word can confuse one person and inform another. The difference is not simply the sound or glyph on a page, but what each person takes them to mean. A large language model (LLM) is trained on large amounts of text to predict likely sequences of words. Because it learns from patterns in linguistic form, it raises the question of whether learning those patterns can amount to meaning. In this essay, I begin with Friedrich Nietzsche’s “On Truth and Lies in a Nonmoral Sense,” which describes concepts as metaphors shaped by abstraction and social convention (Nietzsche, 1873/2006, pp. 115-118). Then, I examine computational linguists Emily Bender and Alexander Koller’s “Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data,” which argues that a system trained only on linguistic form cannot learn meaning. Their argument presents the strongest challenge to my position. Finally, I consider the cognitive science study by Siqi Ning et al., “On Language and Thought: Bilingual Experience Influences Semantic Associations,” which finds that bilingual experience can influence how people judge relationships between concepts (Ning et al., 2020). Together, these texts help explain how linguistic form can shape meaning, while raising the question of whether language alone can provide understanding. I argue that language models do acquire meaning, but of a limited, relational kind. They learn relationships and conventions preserved in text without gaining human-like understanding grounded in sensory experience.

Nietzsche: How Human Concepts Take Shape

If human concepts are shaped by sensory experience, metaphor, and shared linguistic conventions, what remains missing when a model learns from human language alone? Nietzsche’s account of metaphor and convention offers a way to examine that question. Nietzsche describes the formation of concepts as developing through metaphorical steps (Nietzsche, 1873/2006, p. 116): “a nerve stimulus is transferred into an image: first metaphor. The image, in turn, is imitated in a sound: second metaphor.” He also writes that “a word becomes a concept insofar as it simultaneously has to fit countless more or less similar cases” (Nietzsche, 1873/2006, p. 117). A word becomes a concept when we apply it to many cases that resemble one another without being exactly alike (Nietzsche, 1873/2006, p. 117). Nietzsche illustrates this with the concept of a leaf (Nietzsche, 1873/2006, p. 117). Each leaf differs from the others, but we form the general concept by discarding those individual differences. In the same way, he says, humans “dissolve an image into a concept” (Nietzsche, 1873/2006, p. 118). Concepts help us organize experience, but they do so by simplifying it and leaving out some of what makes each case unique.

Nietzsche’s discussion of different languages also connects meaning to social convention. He writes that “[t]he various languages placed side by side show that with words it is never a question of truth, never a question of adequate expression; otherwise, there would not be so many languages” (Nietzsche, 1873/2006, p. 115). Nietzsche links convention to social life. A community establishes shared designations as part of a “peace treaty,” then treats those agreed labels as standards for truth and falsehood (Nietzsche, 1873/2006, p. 115). I take this to mean that no language gives a perfectly adequate expression of reality. Communities use shared words and conventions to organize and communicate their experiences. This does not necessarily mean that different languages attach different labels to one identical meaning. Rather, it suggests that shared meanings are shaped by the conventions of particular languages and communities. Since LLMs are trained on texts in many languages, they encounter relationships among words that reflect a wider range of cultural, historical, and situational contexts.

This account of concept formation invites a comparison with LLMs. A model learns patterns among words from many examples of how they are used. For instance, car is more closely associated with road than with sky. As the model encounters more relationships among words, it can form more reliable associations: fly and sky may lead to plane rather than car. These patterns resemble Nietzsche’s account of concepts fitting many cases, because the model learns how words relate across a large number of examples. The analogy is not exact, Nietzsche is describing how people abstract concepts from experience, while the model learns linguistic patterns from texts. Still, those patterns can allow the model to use words in ways that reflect the concepts and conventions found in human language.

The difference between human perception and a text-trained model remains important. For a person, a car can involve the smell of gas, the sound of its engine, and the sight of its wheels moving. A text-trained model does not sense these things directly, but it can encounter verbal descriptions of them in text, image captions, video transcripts, and stories. I interpret Nietzsche’s account as involving three spheres: the physiological or sensory, the perceptual or mental, and the linguistic. Nietzsche describes each transition as a leap into a different sphere (Nietzsche, 1873/2006, p. 116). Likewise, Peter Bornedal discusses the sequence as nerve stimulus, image, and word (Bornedal, 2010, pp. 55-58). A text-trained model has direct access to the linguistic sphere, while information about sensory experience and mental imagery reaches it through words. I argue that language can still convey information about those other spheres. For example, one description of jumping into the ocean might explain the salty taste and the cold water of the Pacific, while another describes the warm water and calm waves of the South Atlantic. These descriptions present different experiences, and together they enrich the model’s network of relationships around ocean. The model does not have the human experience of entering the water, but the varied descriptions can give it a limited form of meaning. If each person experiences a leaf or an ocean differently, language can still describe similarities among those experiences. In this sense, a rich linguistic network can represent relationships to the sensory and mental worlds without giving the model direct access to them.

Bender and Koller: Why Linguistic Form May Not Be Enough

Bender and Koller argue that “the language modeling task, because it only uses form as training data, cannot in principle lead to learning of meaning” (2020, p. 5186). They define meaning as a relation between linguistic form and communicative intent: what a speaker intends to do or communicate by using an expression. They also recognize that LLMs learn information about language. For example, their review of “BERTology” describes models encoding subject-verb agreement, dependency structures, semantic roles, and lexical similarity. Their point is that these patterns are useful reflections of meaning, but they do not, by themselves, establish a connection to communicative intent or the world (Bender & Koller, 2020, pp. 5186-5187). Their definition raises a question for my argument: must meaning include understanding a speaker’s particular intent, or can a model have a more limited grasp of meaning through relationships among words and concepts? For example, I can know what sadness means without feeling sad at that moment, and I may understand the word without fully understanding another person’s experience of sadness. This suggests a difference between knowing a concept’s conventional meaning and sharing or interpreting someone’s personal experience. Bender and Koller make the relation to communicative intent central. I argue that this is too narrow if it rules out every form of meaning that does not involve human-like understanding.

Bender and Koller also invoke the symbol-grounding problem: dictionary definitions alone cannot teach a person who does not know Chinese what Chinese words mean, because the definitions only lead to other words (2020, p. 5188). But a LLM’s training includes far more than isolated definitions. It includes examples, explanations, descriptions, instructions, and stories showing how words are used in different situations. This connects with my reading of Nietzsche. Nietzsche describes concepts as formed by grouping together many cases while overlooking their differences. The concept leaf, for instance, includes leaves of many colors, shapes, and kinds. A model can encounter those variations alongside descriptions of a leaf’s relation to a tree, seasons, and growth. Bender and Koller anticipate the objection that enough text might solve the problem, and reply that quantity alone is not grounding (2020, p. 5193). My counterargument is that varied human language is more than a larger pile of definitions. It preserves traces of people using concepts in socially shared, world-related situations. I argue that a model can learn a limited form of meaning from those relationships, even if that meaning is not grounded in the same way as a person’s.

Finally, Bender and Koller appeal to human language acquisition, particularly interaction and joint attention. They cite research suggesting that children learn labels more effectively when they and a caregiver attend to the same object than when they merely hear a label from someone behind a screen (2020, p. 5191). This is evidence that children learn language through more than exposure to word forms. However, it does not automatically show that a machine must learn meaning through the same process. A text-trained model does not share a child’s sensory experience or joint attention, but during training it does learn from many examples and contexts rather than from a single definition. This is not equivalent to human interaction. My claim is that it may provide a different route to a limited kind of meaning. Whether that counts as meaning depends on whether meaning must include human-like communicative intent or can also include stable, context-sensitive relationships among words and concepts.

Comparative Analysis: What Distinguishes Human Meaning from AI Language?

The guiding question is whether, if language helps shape human concepts and associations, a model’s reliance on language alone rules out all meaning or only human-like, experience-grounded meaning. Nietzsche and Bender and Koller both challenge the idea that words transparently reproduce reality. Nietzsche describes concepts as metaphorical constructions: our words do not give us direct access to things as they are in themselves (Nietzsche, 1873/2006, p. 117). Bender and Koller distinguish linguistic form from communicative intent: the speaker’s intended meaning in a particular context (Bender & Koller, 2020, pp. 5185-5186). Intent is not simply sensory reality; it can concern the physical, social, abstract, or hypothetical world. Their concern is that linguistic form alone does not give a model the relation between an expression and a speaker’s intent. Thus, both accounts reject language as a transparent label, although Bender and Koller set a further requirement: meaning must connect language to something beyond linguistic form.

Another distinction concerns how humans and models encounter language. Humans experience language through perception and social interaction, while a text-trained model learns patterns from descriptions produced within those human practices. It does not have the same direct experiences or participate in conventions in the same way, so it encounters them indirectly. Ning et al. (2020) found that bilingual experience can influence people’s semantic associations, but their results do not show that bilingual associations are universally better or more accurate. Nor do they prove that multilingual models understand meaning. As an illustration of a possible sound-based link, not an example tested by Ning et al., English fly and sky rhyme, while their French translations, voler and ciel, do not. The rhyme gives English speakers an additional sound-based connection between the words, though it does not by itself show that their meanings are closer. The study does support the idea that linguistic experience can shape human associations. A model trained on texts from different languages, cultures, and historical contexts may develop a broader network of associations, though a larger network does not automatically mean more accurate meaning or prove that meaning is present. Therefore, I think Bender and Koller are right that language models do not thereby gain human-like, experience-grounded understanding. But the absence of that human form of grounding does not, by itself, prove that models have no limited relational meaning at all.

These texts also differ in the kinds of evidence they use, which affects what each can establish. Nietzsche develops a philosophical and rhetorical argument (Nietzsche, 1873/2006, pp. 115-118). His images of metaphor, the leaf, and the worn coin make an abstract claim vivid: concepts are human constructions that become familiar enough to seem natural. This method helps expose assumptions about language, but it does not test how a particular computational system processes words. Bender and Koller use a more analytical strategy. They define the terms “form,” “meaning,” and “understanding,” review language-model research, and use thought experiments to argue that success with linguistic patterns does not demonstrate communicative intent or grounding (Bender & Koller, 2020). Their definitions make the disagreement precise, though their conclusion depends partly on treating intent as a necessary condition for meaning. Ning et al., by contrast, use behavioral judgments and an event-related brain-potential measure across three experiments. Their results provide empirical evidence that bilingual experience can affect perceived semantic associations, while also showing that the effect varies with proficiency and whether participants see pictures or words (Ning et al., 2020). This difference matters to my argument. Nietzsche offers a framework for thinking about convention, Bender and Koller challenge the inference from form to understanding, and Ning et al. demonstrate that linguistic experience can alter associations in human cognition. The study cannot establish that language models have meaning, because human bilinguals have bodies, intentions, and social histories that models lack. It does, however, make it harder to claim that linguistic form is irrelevant to meaning. Taken together, the texts support a qualified conclusion: associations learned from language may count as a limited kind of meaning, but evidence of association alone does not show that a model has human understanding.

Counterargument and Limits

An objection to my argument is that an LLM may reproduce associations without having the experiences those associations describe. Nietzsche’s metaphorical account moves across different spheres (Nietzsche, 1873/2006, pp. 116-118): from a physiological or sensory stimulus, to a mental or perceptual image, and then to a linguistic sound or word. Imagine I tell my child that falling in love can bring passion, joy, excitement, and constant thoughts of another person. My child can understand the words and repeat my description, but may not yet grasp what they mean as a lived experience. Years later, after falling in love for the first time, those words may connect to emotions and bodily sensations the child has now experienced. This illustrates Nietzsche’s distinction: language can describe and organize an experience, but it does not reproduce the sensory and mental experience itself. An LLM can make similar associations from descriptions in text, yet it has not undergone those experiences. The same difference appears in the example of skydiving: reading that someone calls it “the greatest adrenaline rush” can convey ideas of height, fear, and exhilaration, but it is not the experience of jumping from a plane. A model can identify the claim as subjective and describe how people use such language, but it cannot verify what the event felt like to that individual.

This objection identifies what is missing from the model’s access to the spheres: it learns about sensory experience and mental imagery through linguistic descriptions, rather than reaching those spheres in the human way. But that may limit the kind of meaning it has without proving it has none. People can understand concepts and descriptions of experiences they have never personally had. Therefore, I argue that a model may develop a limited linguistic or relational meaning while lacking the experience-based meaning a human can have.

Conclusion: What Language Can Teach Us About AI

To conclude, my analysis of Nietzsche, Bender and Koller, and Ning et al. shows that meaning is difficult to define precisely. Words, phrases, and sentences gain significance through their relations to other expressions and to shared conventions, creating layers of interpretation rather than giving us direct access to the things they describe. Nietzsche emphasizes the metaphorical and conventional character of language. Bender and Koller distinguish linguistic form from communicative intent. Ning et al. show that bilingual experience can influence people’s semantic associations. Together, these accounts suggest that human meaning is also shaped by perception, language, and culture. Bender and Koller are right that linguistic patterns alone do not guarantee that a model can recover a speaker’s intent. Yet, my analysis suggests that LLMs can form associations from the patterns and conventions preserved in text. A model can describe love or skydiving without having lived those experiences. Words can convey many things about these phenomena, but they cannot give the model the first-person experience of them. This limits a model’s understanding, but does not prove that it has no meaning at all. A text-trained model may acquire a limited, relational form of meaning without human sensory experience. The remaining question is not whether a model’s language perfectly captures human experience, but whether a limited, text-based network of relations deserves to be called meaning. My answer is yes, provided that this claim is kept distinct from the stronger claim that a model experiences or understands the world as a human does.

References

• Bender, E. M., & Koller, A. (2020). Climbing towards NLU: On meaning, form, and understanding in the age of data. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (pp. 5185-5198). Association for Computational Linguistics. https://doi.org/10.18653/v1/2020.acl-main.463
• Bornedal, P. (2010). The surface and the abyss: Nietzsche as philosopher of mind and knowledge. Walter de Gruyter.
• Nietzsche, F. (2006). On truth and lies in a nonmoral sense (D. Breazeale, Trans.). In K. Ansell-Pearson & D. Large (Eds.), The Nietzsche reader (pp. 114-123). Blackwell Publishing. (Original work written 1873)
• Ning, S., Hayakawa, S., Bartolotti, J., & Marian, V. (2020). On language and thought: Bilingual experience influences semantic associations. Journal of Neurolinguistics, 56, 100932. https://doi.org/10.1016/j.jneuroling.2020.100932