Given that magic eye pictures are basically a certain convolution of a repetitive source image (or random noise) and a depth map, it ought to be possible for an AI, if not a standard procedural algorithm, to at least attempt to “factor out” the source and map from the result image.

Whether that even counts as seeing and whether this guy’s neural network could do this, or maybe “see” in some other way is open to question. But I’m guessing it’s a no.