> don't necessarily generalize to the whole population

In my experience it is far more likely to mean "doesn't generalize well, and moving to a reasonable training set will reveal flaws in the model.