I think it means dont under or overfit.

How well your model can predict unseen data, data it was not trained on.

It means that let's we trained our model for image classification task with accuracy of 90 percent for three class like negative ,positive , none then we have test data or new images of these classes on which our model is not trained it also can predict with almost equal to accuracy of class as it was on train data. for example on new accuracy will be around 89 to 91 It means that our model is generallizing well. It does not mean that we introudce new class to our model it can predict it accuratly.