I totally agree. It's not the first time I read how DLSS is superior(better image) than native, but that's just not logical. If someone perceives It like that, then It's most likely because post-processing for native is just bad.
I wonder If developers don't do It intentionally, so everyone starts using upscalers, which would save them time optimizing their games.
It's due to camera jittering. The idea (common to DLSS too, and maybe other TAAs) is to apply sub-pixel translations to the camera each frame and then accumulate the results.
It's similar to MSAA, but in this case each pixel is sampled multiple times each frame hence the performance cost, check the pattern at
https://www.guru3d.com/articles-pages/radeon-6950-6970-review,7.html
In DLSS/FSR each pixel is only sampled once per frame (so no immediate performance cost), but the sampling position changes each frame. Now, the really interesting question is: how many samples are accumulated per pixel in FSR2/DLSS?
The answer depends on quality mode, but in the highest quality mode it's 18 samples, and it goes all the way up to 72 for ultra performance mode. See
https://github.com/GPUOpen-Effects/FidelityFX-FSR2/blob/master/README.md#camera-jitter
So after 18 frames, how many samples are accumulated on screen? For a 4K output:
Native: 3840x2160x1 = 8.294.400
MSAA 4x: 3840x2160x4 = 33.177.600
FSR Quality: 2560x1440x18 = 66.355.200
So it's not only about post-processing, there IS more information in each pixel when using DLSS/FSR, although you have to wait for it.
When things move on screen, both algorithms use the motion vectors generated by the game engine for each pixel to continue the accumulation process using the color of the respective pixel from the previous frame.
Quality suffers momentarily, but it's better than dropping down to 1 sample per pixel. So how long does it take for FSR to match the number of samples of native rendering? It's 3 frames for native and 9 frames for MSAAx4, so it's pretty fast.