How do LLMs understand words?

Language on the internet changes faster than any textbook can keep up. A word is no longer just a term with a neat line in a dictionary. It can be a caption that signals sincerity one day and satire the next. It can be a meme that travels across platforms and returns with a different tone. Into this swirl step large language models, which answer us with fluid sentences and the confidence of a well briefed colleague. We ask how these systems understand words. The honest answer begins with a correction. They do not understand in the way people do. They predict. They operate by learning how words tend to appear together, which other words cluster near them, and what patterns usually follow a given phrase. The result looks like comprehension because we often measure understanding by how well someone continues a conversation. A model excels at continuation because continuation is exactly what it was trained to do.

To picture this, imagine language as a crowded city. You do not need to memorize every street to reach the coffee shop with the red stool by the window. You can follow the paths you have taken before, note the landmarks, and infer the rest. A language model moves through this city in a similar fashion. During training, it reads enormous amounts of text and learns the routes that people take with words. It learns that technical discussions contain certain signposts. It learns that a phrase like low key often precedes a soft confession or a wink. It does not possess a private sense of meaning. It recognizes typical neighborhoods for words and the directions in which sentences usually travel.

Slang exposes both the strength and the limit of this approach. Consider how a word such as delulu moves from a fandom niche to wider social platforms, and then into brand copy. Each hop shifts the tone. A model can follow the shift because it has seen thousands of examples across posts, comments, replies, and articles that explain the joke for those who arrived late. What looks like cultural insight is a careful alignment with the paths the word has taken before. The model can reproduce the next step in the dance because it has watched the dance many times. If you bring it into a new room where the rules differ, it falters. A term that carries one meaning on a gaming forum may signal something else in a financial chat. The system does not ask a friend to clarify. It averages across rooms that resemble this one and tries to infer the local rhythm.

It can help to think about gravity. In this city of language, certain words pull others closer. Policy attracts implications and framework. Messy draws in authentic and algorithm and feed. The pull is statistical rather than moral. That is why the same model that sounds expert in a familiar setting can be decisively wrong in an unfamiliar one. It is reading the sky of words rather than the ground truth of events. A human speaker can bring memory, emotion, bodily context, and lived stakes to a sentence. A model brings a map of co occurrences, and the talent of producing the next likely token in a chain.

The internet complicates this in two ways. It makes the job easier because people leave rich trails. We write captions and comments, send emails, publish newsletters, debate in forums, and post micro essays while waiting for the train. All of that becomes training material that teaches the model how language moves. At the same time, the internet makes the job harder because so much of our language is performance. We say things for the bit. We speak in irony that is legible only within the bonds of a particular group of friends. We type in lowercase to signal softness. A model can learn that lowercase often accompanies intimacy. It cannot know that you cried before you pressed share. It can learn the shape of the ritual. It cannot witness the moment that gave the ritual its weight.

Data carries power, and that power shows up in output. The language a model returns is the language it could collect at scale. Voices that travel far are easier to learn. Dialects that platforms standardize through autocorrect become overrepresented. A meme born on a Manila side street may reach a global audience through the gloss of a brand voice, and the model will mostly learn the polished version. When you type a word the way your aunt says it, the model may nudge it toward the standardized spelling. It believes it is being helpful because the probabilities point that way. It cannot ask whether it has erased a detail that matters.

We can watch the same process play out in smaller arenas that feel familiar. Dating app prompts fill with one liners that worked last week. People borrow and refine those lines to match what they think the app rewards. Then they ask a model for a better opener. The result resonates because the neighborhoods of flirtation are densely mapped. Many of us have walked those streets, so the model can light the way. In offices, Slack creates another kind of ritual. Phrases like quick sync carry a script that everyone recognizes without being told. The model learns to produce follow ups, summaries, and polite sign offs that fit the pattern. It is not empathizing. It is performing choreography that succeeds precisely because it repeats across companies and time zones.

From time to time the mask slips. The model misses a joke that pivots on a small cultural hinge. It offers a factual claim where a friend would offer a pause. It asserts certainty where tenderness would be more appropriate. Those moments reveal the boundary between a predictive system and a person with history, pain, and care. The model cannot sit with you in the space between a question and an answer. It can only accelerate toward what usually comes next. That acceleration feels like intelligence when the stakes are low and the topic is familiar. It can feel hollow when the moment calls for judgment, restraint, or a memory of who said what at dinner last year and why it mattered.

If this is prediction rather than human understanding, why do we keep returning to it? Because prediction has practical value. A system that can continue a sentence in a way that most readers accept as coherent is useful for many tasks. It drafts, summarizes, translates, and adapts tone. It can act as a mirror that does not tire. It can serve as a coauthor that never asks for a weekend. It is autocomplete, extended from a handful of words to the scale of an essay. When a sentence lands, it is not because the model found hidden truth inside a word. It is because the model stood in the part of the city where our conversations tend to flow and pointed in that direction.

The question about understanding also reflects our own habits. We often reward fluency over depth. We like responses that feel familiar and confident. We accept prediction as meaning because meaning is hard work and prediction is fast. In school, a dictionary settled arguments. Online, a word proves itself by how it moves through communities. The argument lives in the comments and the remix. The model captures that motion and returns an average of it. It does not decide what a word should be. It records how the word was used yesterday across many rooms at once, and then it builds a likely today from those yesterdays.

Recognizing this does not require cynicism. It invites care. We can ask what kinds of language we want to amplify through our own use. We can notice when a model smooths out the rough edges of a dialect that matters to us. We can demand that systems include sources beyond what travels most easily. We can measure success not only by fluency but also by humility. A model is at its best when it helps us write clearly, think aloud, and surface patterns we would otherwise miss. It is at its worst when we ask it to substitute for experience, or when we treat probability as wisdom.

So how do language models understand words? They map usage, not belief. They learn which neighbors a word keeps and which streets follow certain turns. They smell the air for pattern rather than meaning. They continue the dance because they have watched so many dances before. This is not a soul that knows a truth. It is a planner that learned our shortcuts and offers new routes that resemble the old ones. Sometimes that is exactly what we need. Sometimes it is not enough. The challenge for us is to know the difference, and to bring our own judgment to the places where prediction falls short.