That makes more sense if the gradient is radial. For a linear gradient, inversion gives the appearance of a changing light source - just like the example given shows.
A radial gradient, on the other hand, might more typically assume an 'oveerhead' light, which is why inversion makes sense.
For me the linear gradient hints that the surface is curved on one axis, i.e. is cylindrical, not spherical.
Of course the context plays a role too on the perceived illusion; the appearance of changing light source may be stronger for lonely button (vs. group of elements sharing the same light source).
A radial gradient, on the other hand, might more typically assume an 'oveerhead' light, which is why inversion makes sense.