Yea, the jaggies are 'somewhat' new...
Thing is, for regular texturing, MSAA was developed as a good way to handle texture aliasing at polygon edges. The rest of the polygon was antialiased by bilinear/trilinear/anisotropic filtering.
But then advanced shading methods appeared, like bumpmapping, displacement mapping, horizon mapping etc. And guess what? MSAA doesn't handle those, neither do the conventional texture filtering techniques. Why not? Because those 'bumps' and things aren't just on polygon edges, so MSAA doesn't handle them. And the bumps aren't simple linear functions, so conventional filtering does not predict the aliasing, since that filtering assumes flat polygon surfaces.
So per-pixel lighting tricks are very susceptible to aliasing anyway. And yes, deferred rendering techniques make it even more difficult, because the MSAA and other geometry info get lost between passes.
SSAA isn't the only way, but it certainly is a good way to solve it.
Other approaches include encoding extra data into your normalmaps and/or other textures, to avoid aliasing during shading itself.
I recall 3DMark05 using an interesting trick where they had 'denormalized' normalmaps. As in: the normals that were stored, were not necessarily unit length. The variation in length would act as a scaling factor during shading. When the normalmap was sampled via a bilinear filter, the length was 'averaged', resulting in somewhat smoother shading.
Which was fun, when ATi tried to claim that 3DMark05 was being unfair to them because they didn't use 3DC texture compression. But 3DC is designed to always compress/decompress unit vectors. So it was entirely incompatible with this antialiasing trick.