The one idea
AI read most of the text humans have written: articles, books, transcripts, forum posts. It broke all of that into small pieces called tokens, and learned which token tends to come next. When you type something, it predicts the most likely next token, then the next, and builds an answer that way.
Say "twinkle twinkle little" and it returns "star." Not because it knows the song, but because that is what almost always comes next. That is the whole trick.
What that means for you
- It is not magic, it is pattern matching. A useful way to hold it: pattern recognition in a fancy suit. It is not thinking, it is predicting.
- Garbage in, garbage out. Give it a vague one-liner and you get a vague, average answer. Give it the full story and a clear ask and the answer gets sharp. The output never beats the input.
- It mirrors us. The repeated quirks you see in AI writing are it copying how humans write, not showing a preference. Same reason it can sound confident while being wrong: confident text was common in its training.
The practical takeaway
Treat it like a sharp assistant who knows a lot but knows nothing about your situation until you tell them. The fix is context. The structured way to feed it context is the 4-part prompt formula, the next resource below.