Easy to train the AI if the camera is fixed and the test is repeatable with the conditions very similar to how they are ran every time.

It's like using a baseball launching machine in an enclosed space to hit a non-moving target. You can make it as accurate as possible in a lab test (benchmarks) but it would never be as good in real-world conditions (i.e. actual gaming).

Eventually, using DLSS results would be considered as bad as the smartphones cheating on the benchmarks.