It doesn't blur the image as such - the data just isn't there to not be blurry - the current implementation seems to be hugely behind temporal upscaling approaches never mind AI/DL driven.
The problem is, there simply is no other solution. You either use temporal accumulation to gain more information, or you use deep learning to create fake but statistically pleasing artificial information ( or both, whixh is what DLSS 2 is). Otherwise you are left with the same linear techniques from the 1970s, e.g. bicubic with some sharpening. If you integrate earlier you can use depth buffer to better treat edged,but nothing yéu csn do about texture and shader detail