What makes AI 'generative'?

Your photo app looks at a picture and says “beach”. That’s impressive — but notice its answer is one word chosen from a fixed list. It recognises. Now compare an image tool that, given the words “quiet beach at dawn, watercolour”, paints a picture that has never existed. It generates.

From labelling to producing

Everything you met in lessons 1–4 was recognition: spam or not spam, ripe or not ripe, cat or dog. The model’s output is a label or a score.

Generative AI turns the machinery around. Instead of asking “which label fits this input?”, it asks, over and over: “given everything so far, what small piece most plausibly comes next?” A next word. A next patch of pixels. A next slice of audio. String thousands of those small predictions together and you get a paragraph, an image, a voice.

Why “plausible” is the key word

The model was trained on enormous amounts of existing text and images (lesson 3: patterns from examples). So each small prediction is really: “what would typically follow here, in material like my training data?”

That single sentence explains most of what you’ll experience with these tools:

Why outputs feel fluent and human-like — human material is what the patterns came from.
Why outputs can be generic — “typical” is the default.
Why outputs can be wrong while sounding right — plausible and true are different properties. Lesson 7 is entirely about this.

Try it now (2 minutes)

Sort these into recognise or generate: unlocking your phone with your face · a chat assistant writing a birthday message · translating a menu photo · an app inventing a recipe from your leftovers · flagging a fraudulent card payment.

Check your understanding

1. Which of these is generative AI? Marking an email as spam Writing a new cover letter from your notes Detecting a face to unlock a phone

2. A generative model chooses each next piece by asking: What is objectively true here? What most plausibly comes next, given patterns from training? What did the user secretly want?

Recap

Recognisers label; generators produce, one predicted piece at a time, guided by plausibility rather than truth. Next: the specific generator you’ll use most — the large language model — and what it’s really doing when it “talks”.