Why an AI understands stick figures

A sketch is astonishingly little information: a few strokes, no colour gradient, no texture. And yet humans and AI alike instantly see a house or a cat in it. Why does this work so well? An explanation without formulas.

Recognising means finding what's typical

We don't recognise things by comparing every detail, but by re-recognising their characteristic shape. A house is “a square with a triangle on top”, a cat is “a round head with pointed ears”. These reduced basic patterns are often called prototypes. A good sketch hits exactly this prototype – and that's exactly what an AI looks for too.

How an AI learns this

During training, a neural network is given millions of example drawings, each with the correct label. Nobody programs rules like “ears = cat” into it. Instead, it adjusts its internal values until it reliably tells the examples apart. In the process a hierarchy of features emerges: early layers respond to simple edges, later ones to corners and curves, the topmost ones to whole shape parts like wheels or roofs.

In the end the network doesn't give a single answer but a probability for each possible category. “80% cat, 10% dog” – those are the bars you see in the game. The technical details are on the About the AI page.

Why less is often more

Surprising to many: a simple, clear sketch is often recognised better than a detailed work of art. The reason: too many strokes add ambiguity. Shading, backgrounds or playful ornaments suddenly resemble other categories and “dilute” the typical pattern. Whoever reduces to the core shape gives the AI an unambiguous signal.

Where the limits are

An AI doesn't “understand” a subject in the human sense – it recognises statistical patterns. That's why it fails at things a human classifies effortlessly: at very atypical depictions, at words it never learned, or at subjects that are formally very similar (cat, tiger and lion all have a round head with ears, after all). Exactly these mix-ups make the game appealing – and show how our own perception works.

Test the theory

Draw a subject minimally first, then with lots of detail – and watch when the AI becomes more confident.