The guessing game
At its heart, an AI learns language by playing a fill-in-the-blank game. Take a real sentence, hide the next word, and make the model guess it. Then show it the real word. If the guess was wrong, nudge the model a little so it does better next time.
That is the whole idea. It sounds almost too simple to work. The magic is in the scale. The model plays this game across a huge amount of text, making an unimaginable number of guesses and tiny corrections. Bit by bit, its guesses get better, and to get better it has to pick up how language actually works.
Why guessing words teaches so much
Here is the clever part. To guess the next word well, the model is forced to learn all sorts of things it was never directly taught. Think about what it takes to fill in these blanks:
“She poured the milk into her...” → to guess “coffee”, it learns how everyday objects go together.
“The capital of Japan is...” → to guess “Tokyo”, it picks up facts.
“He was so tired that he...” → to guess “fell asleep”, it learns cause and effect.
Grammar, facts, tone, cause and effect: all of it gets pulled in as a side effect of trying to guess the next word. The model never studies these things directly. It absorbs them because they help it win the game.
It learns grammar like a child, not a student
Notice that the model never reads a grammar book. It does not learn the rule for where a comma goes or how to make a verb agree with its subject. It learns the way a small child does: by hearing correct sentences again and again until the right pattern just feels natural.
This is why AI writing is usually smooth and grammatical, but also why it can be confidently wrong. It learned what correct language looks like, not what is true. Good grammar and true facts are different things, and the game only trains one of them directly.
How it handles many languages
If the text a model reads includes many languages, it learns them all the same way, by guessing the next word in each. That is how one chatbot can answer in English, Spanish, or French, and even translate between them. It simply learned the patterns of each language it saw. It is usually strongest in the languages that appeared most in its reading, and weaker in ones it saw little of.
Does it really know what words mean?
Sort of, but not like you. Through all that practice, the model learns that some words show up in similar situations. It learns that “king” and “queen” behave alike, and that “hot” and “cold” are opposites that appear in similar spots. So it acts as if it understands meaning.
But there is no real-world experience behind the words. It has never felt heat or seen a king. It has a powerful sense of how words are used, built entirely from text. That is a real and useful kind of knowledge, and it is also why AI can sound like it understands far more than it truly does.