It creates leaf nodes that perfectly memorise training data, including noise

It always produces linear decision boundaries

It requires normalised features to function correctly

It uses too few features at each split