How an AI model is trained
Where the 'knowledge' comes from, in three plain stages — and why the cut-off matters.
Updated 20 May 2026
A model isn’t programmed with facts the way a calculator is. It’s trained — shown enormous amounts of text and adjusted, over and over, until it gets good at predicting what comes next. Understanding the three rough stages explains a lot of its quirks.
1. Pre-training
The model reads a vast slice of the public internet, books, and other text. It isn’t memorising; it’s tuning billions of internal numbers so its next-word guesses match real writing. This is where the broad “knowledge” and language ability come from — and it’s frozen at a point in time, the model’s training cut-off. Ask about something newer and it simply wasn’t there to learn from.
2. Fine-tuning
Raw pre-training produces something fluent but unfocused. Fine-tuning teaches it to behave like a helpful assistant — to follow instructions, answer in a useful shape, and refuse genuinely harmful requests.
3. Alignment from human feedback
People rate competing answers, and the model is nudged toward the ones humans prefer. This is what makes a modern assistant feel polite and on-task — and also why two models can have noticeably different “personalities”.
The training cut-off is why an assistant can be brilliant on timeless topics and wrong on last week’s news. When recency matters, check a live source.