I suppose that’s more because it’s not doing a full physics simulation it’s just parroting the cloak-specific physics it observed but still

This. I'm sure to a sufficiently intelligent observer it would still look wrong. You could probably achieve the same thing with a conventional algorithm, it's just that we haven't come up with a way to profitably exploit our limited perception quite as well as the ML does.

In the same vein, one of the big things I'm waiting on is somebody making a NN pixel shader. Even a modest network can achieve a photorealistic look very easily.