These three forms are trained in order to approximate the ground truth, and they perform significantly better than their non-neural counterparts which do the same thing but worse. Moreover, neither of them is a pure postprocess, since they receive various rendering passes as input, like the depth buffer. They don't really hallucinate anything, just interpolate.
This is another story with DLSS 5, which doesn't reconstruct some ground truth and which is a pure post-processing effect. It just makes the final image look more photo-like in a vague sense. This is more of a "slop filter".