The "lost precision" is the difference between this and this. And that's just the two-dimensional case. Consider it in three dimensions, especially with perspective projection which means objects far away will need more samples by definition since you see more of the scene per pixel. Have you ever played vanilla minecraft? Move the mouse one pixel to the left, and the whole screen seems to jitter randomly, because at a distance, much more than one texel (texture pixel) of information is condensed into a single pixel, and depending on the view angle, any one texel is chosen randomly, which produces a weird color soup which is very straining on the eyes - that's what the lack of multisampling results in. If you used multisampling, you could weigh each texel according to its contribution to the pixel's color, and produce a correct pixel color which will not change just by looking at it from a slightly different angle. It adds stability to the render.
If a pixel is made 75% of one information source and 25% another, it would usually just use the 75%
You just lose some precision
(though note that while ray-tracing solves this naturally by jittering the camera ray by random amounts to capture subpixel information, traditional graphics hardware does this by mipmapping, i.e. using less detailed textures at a distance, which is cheaper than full-on multisampling. unfortunately minecraft uses neither, but that's just an example).