Like the nurture versus nature debate in biology, linguistics grapples with two opposing views: prescriptivism and descriptivism. These perspectives explore whether language should be governed by strict rules or understood in its natural use. Prescriptivists view language as a set of rules that dictate how it should be used. Descriptivists observe language in its natural use and aim to describe it as it is. Who is right seems to be a futile battle and of course, a reasonable person may take a hybrid approach, seeing value in both. Yet, the dawn of large language models supports the primacy of the descriptivist outlook.
Prescriptivism is a model that restricts the fluidity and creativity of language. Without flexibility, language becomes rigid and loses its ability to evolve.
Descriptivists mirror the way humans actually use language, capturing its fluidity and creativity.
The Problem with Prescriptivism
Your middle and high school English teachers were almost definitely prescriptivists and with good reason. They were trying to teach you the standard English that will help you excel in business and academic settings. However, for these same reasons, prescriptivism also acts as a gatekeeping device, discriminating based on class and educational attainment. It is a tool of traditional power structures. Many of these more arbitrary rules, it has been argued were designed to force English to follow patterns of Latin as a show of status. For example, the prescribed rule against split infinitives is a favorite example. An infinitive is the base form of a verb, often preceded by “to,” such as “to run” or “to write” in English. In Latin, splitting an infinitive (i.e. “to quickly run”) is impossible since it is only one word, but as previously mentioned, two words in English (i.e. “currere” vs “to run”). Hence, not splitting an infinitive reflects a descriptive rule in Latin, where it is structurally impossible, but in English, it becomes a prescriptive rule imposed to mimic Latin conventions.
Unfortunately, prescriptivism is a model that restricts the fluidity and creativity of language. As a set of predetermined and often arbitrary rules, prescriptivism is language without art. In fact, it is also the absence of language itself. Language would not be able to evolve in a purely prescriptivist environment. But of course, who does not want their child to learn the more employable language? So, I do not fault any of your former teachers or parents for their prescriptive focus. I will however argue that as adults, we should use and understand language from a more sophisticated paradigm, a paradigm that is supported by evidence.
Weathermen do not create the weather and lexicographers do not create the language.
Language is Mathematical
The Contradiction of Language as Non-Mathematical but Linear
To transition from prescriptivist constraints, consider how people perceive language—particularly its development and structure. You will hear some interesting insights. Many will claim to be more linguistically minded versus mathematically as if there is some obvious solid demarcation between the two. Clearly, you can only be from one side of this mutually exclusive dyad. The claim clearly implies that language is non-mathematical. They will also often describe language in linear terms, a clear violation of the non-mathematical view of language. One of the most simplistic versions of this comes from my childhood when I learned the Spanish alphabet, an adult stated, “oh, you speak Spanish now!” Of course, I disagreed, but he assured me that all I needed now was to replace each letter with its Spanish equivalent, a simple linear bijective and clearly flawed function on the language. A more complicated linear model of the form y=mx+b would be the subject plus verb plus object equals an indisputable, universal meaning. All elements are found in a comprehensive dictionary with minimal alteration following another set of simplified rules such as verb conjugation. Now, there is some validity to this argument, but it is still oversimplified. Like a weatherman observing and predicting weather, lexicographers observe, record, and describe language without shaping it.
Language as Mathematical and Non-Linear
Building on this contradiction, the linguistic fluency of Large Language Models (LLMs) provides evidence that language is inherently mathematical. These models are mathematical, finding patterns based on probabilities based on context. However, they are largely non-linear. Even the best crafted sentences can induce some ambiguity; we will not all agree on the exact meaning of every sentence.
LLMs adapt dynamically to context, reflecting the adaptability and fluidity inherent in human language.
Meaning in Context
A sentence’s meaning relies on the words used and the context in which it is understood. I hope that we can agree that a sentence’s meaning is at least dependent on the collection of words within that sentence, right? But even each word itself is defined by the context of the sentence in which it exists in real time from the receiver’s perspective. A traditional dictionary cannot capture such complexity. For example, consider the word branch. It can mean a part of a river, a part of a tree, or a bank or a business location. Sure, one could theoretically list all ways this word could be used but would it not make more sense to listen to the sentence or conversation in its entirety and know what it means? What if we want to use it in a more novel or abstract way such as the branches of meandering thought?
Self-attention transformers revolutionized AI by focusing on contextual relationships, much like human comprehension.
Attention Transformers
The landmark paper, Attention is all You Need (Vaswani et al., 2017), introduced self-attention transformers—a revolutionary mechanism in AI that allows models to focus on the relationships between words in context, rather than processing them sequentially. This innovation drastically improved the efficiency and accuracy of language models by enabling them to weigh the importance of each word in relation to others. Yet, if we step back from the technical implications, the cultural implications appear to support this view of words being defined in context. Self-attention transformers do not simply compute relationships between words; they model how meaning emerges dynamically from the context of surrounding words, much like human comprehension.
LLMs are Descriptive Linguists
This progression leads us to LLMs, which can be seen as modern descriptivist linguists. They did not learn by a collection of rules but their language ability emerged from a large neural network that was trained on actual language as it is used by humans. Of course, some rules exist in their coding but that is not their foundation. Their existence shows the limits of prescriptivism when exploring the structure and evolution of language. If rules and dictionary definitions made a language, the problem of computer language would have been solved in the 1970’s or 80’s. Rather, it took the breakthrough of artificial neural networks and attention transformers to truly crack the problem, 40 plus years later. This suggests not only can definitions be defined by context and in real time but also that they must be. Furthermore, LLMs using attention transformers also model the structure of language, or syntax, similarly, through emergent properties implicit in the training corpora (the body of literature on which the model is trained). These emergent properties arise from patterns identified in a vast training dataset, reflecting the adaptability of LLMs to natural language without being explicitly programmed for specific rules. This enables LLMs to generate non-linear sentence structures with embedded clauses and unconventional yet coherent syntax. So, like descriptivists, LLMs do not work by a simplified pre-defined set of rules, adapting dynamically to context, allowing them to produce language that is human-like.
Language thrives and evolves when observed and described rather than rigidly prescribed.
Language as Art
Finally, language mirrors art in its evolution. Just as a painter learns craft and technique to understand what is generally preferred, they eventually evolve to make statements, breaking rules at the right moments to create beauty and novelty. Rules offer initial guidance, but they can stifle growth if not transcended. Visionary artists push beyond technical mastery to innovate and redefine their craft. Similarly, LLMs generate language by adapting dynamically to context and patterns, reflecting the adaptability and fluidity inherent in human language. Although they may not be capable of producing art, they reflect these inherent qualities of language itself. Who offers the language more: the person who can recite and follow a list of rules or the person who can create a novel, intelligible word that the receiver understands with little to no explanation? Language is inherently beautiful and creative—it thrives and evolves when observed and described rather than rigidly prescribed.
For businesses leveraging natural language understanding, context-based systems like LLMs are essential to staying competitive in an increasingly AI-driven landscape.
Implications
Considering these insights, decades of failed attempts at creating a comprehensive rules-based language model demonstrate that this path offers little. For businesses and organizations leveraging natural language understanding and generation, context-based systems like LLMs are essential to staying competitive in an increasingly AI-driven landscape. These systems are not only more effective but also align with how language naturally evolves and operates. If a prescriptivist inspired model of language could encapsulate the complexity of language, one would have been developed already, years before the dawn of LLMs. Hence, this is unlikely forthcoming and adaptive, context-based systems will dominate. Language generation will require specialized tools, such as LLMs, which are adept at exploiting probability and patterns to model the fluidity and emergent properties of language. This shift reflects a growing recognition that practical applications—from automated customer service to advanced analytics—depend on tools that embrace descriptivist principles. Prescriptivism is a valuable tool for learning, providing an initial foundation. Just as the painter must master a proper brushstroke before violating the prescribed methods, language must be unbound so that creativity and fluidity take precedence. Therefore, adopting a descriptivist paradigm is essential for leveraging the full creative and functional potential of language in modern applications, whether in AI systems or human-centered design.
Conclusion
Language is more than a rigid set of rules; it is a living, evolving system that reflects human creativity and complexity. While prescriptivism offers structure and utility, true understanding comes from embracing the fluidity and adaptability that descriptivism recognizes. The advent of LLMs underscores this perspective, showcasing how context and emergent patterns define meaning in ways static rules cannot. By adopting this more sophisticated view, we can better harness the potential of language—in technology, communication, and beyond. Surprisingly, computer scientists and linguists showed that language is emergent from complex and nonlinear beauty.
Comments are welcome and encouraged!
I have so many more thoughts especially around the mathematics of language to discuss but if I keep adding, this will soon be an entire book. So, instead, share your thoughts in the comments sections and we can engage on this some more. Please engage with the ideas presented rather than relying on appeals to tradition or authority. Let’s have a thoughtful discussion about how language actually works and evolves.
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30. https://arxiv.org/abs/1706.03762
You argue that LLMs function as linguists, trained to learn patterns and context in language. But do you think AI and humans truly learn the same way? What are the key differences (or similarities) in how they process and understand language?
That’s a great point! The differences in mechanics—binary approximations on a series of mathematical operations versus neurotransmitters and electrical impulses—are certainly striking. But I find it fascinating that, like babies, LLMs rely on consuming massive amounts of language to progress from random outputs (cooing) to something structured and meaningful (words and then sentences). That similarity makes me wonder: even if the underlying processes are very different, do you think the way that they abstract patterns from data could be seen as analogous? Could this be why LLMs seem so capable at capturing the nuances of human language? I personally do not know.