The specific point is that we are using linear averages (over large collections of objects, such as galaxies) to model a non-linear effects (General Relativity

The specific point is that we are using linear averages (over large collections of objects, such as galaxies) to model a non-linear effects (General Relativity interactions between individual objects). It may be accurate to use those averages, or it may not be if for example the chaotic nature of N-body problems causes reality to diverge significantly from the simpler model; we don't know how large the error actually is and currently don't have a way to measure it.