The model doesn’t know the rules of the shapes. It shrinks your drawing to a 16×16 grid and turns each square into a number. Then it compares those numbers with your examples. Every example gets a vote: the more it looks like your drawing, the stronger its vote. It says the label with the most votes.
This is called the nearest neighbor method. The model is only as good as the examples it has seen: give it too few, or examples that are all alike, and it gets things wrong.
In the Trick it tab you switch roles: you hunt for the model’s weak spots. Real AI teams work like this too. They push their model on purpose and add the examples it gets wrong to the data, with the right label.
Try this: on purpose, add a shape to the wrong class a few times, then test that shape. Wrongly labeled data fools the model too. Also try drawing a shape upside down or sideways: if it’s never seen an example like that, it gets confused.