Lesson 2 of 9 · AI Basics
How AI is trained
A chatbot doesn't come out of the box knowing things. It gets there in four stages, and knowing them explains a lot of AI's odd behavior.
In the last lesson you saw that AI answers by predicting the next word. The obvious question: how does a prediction machine get good enough to write a legal letter or debug a spreadsheet formula? The answer is training.
Tap through: how a model gets made
1. Read
The model is fed a huge library of text: websites, books, articles, code. It doesn't memorize pages like a filing cabinet. It absorbs patterns in how words follow each other.
The part most people miss
Notice stage 4: by default, the model is frozen when you chat with it. If you tell ChatGPT your dog's name today, the underlying model has not "learned" it. Apps bolt on a memory feature that saves notes between chats, but that is the app taking notes, not the model retraining itself.
This is also why a model can be out of date. Its reading happened before a cutoff date. Anything that happened after that, it only knows if the app looks it up on the web for you.
Why models say "as an AI..."
Stage 3, the coaching, is where a model picks up its manners: declining harmful requests, admitting uncertainty, keeping a helpful tone. When a chatbot politely refuses something, that is not a law of nature. It is the coaching its maker gave it, and different companies coach differently. That is one reason the same question can get different answers from different chatbots, which is exactly where the next lesson picks up.
Last updated August 11, 2026