I think they mean "always (statistically) significantly larger". They're probably imagining that something like cross-validation would make test errors approximately equal to training errors, but if you consistently see significantly larger errors, then you've overfit.