Generative AI: how it creates content

Traditional AI classifies: is it an invoice or a receipt? Generative AI creates: a new email, an image, or a code snippet.

How does it do this? It has processed vast amounts of data and learned patterns in language. If you type “Dear customer, we regret that …”, it guesses “the delivery is delayed” or “an error has occurred” — phrases that often follow. It builds a complete text one word at a time.

It guesses; it does not verify. Therefore, the text can be fluent and professional — yet factually incorrect.

Image generation works on the same logic, but with pixels: from random noise, it builds an image step by step that matches your description. Six fingers on a hand? The model has not learned that hands have exactly five fingers — it has only learned how hands often look in images.

Three key points to remember:

1. It generates new content; it does not copy (though it builds on what it was trained on).
2. It can be incorrect even if the text or image looks credible.
3. It has no understanding of truth or falsehood; it optimises for probability.