I think it actually shows this kind of image recognition AI doesn't actually think, i.e. it hasn't developed high-level concepts of the things in the pictures i

I think it actually shows this kind of image recognition AI doesn't actually think, i.e. it hasn't developed high-level concepts of the things in the pictures it's trained on. When these neural networks get an input image and spit out labels like "bird" or "car", they haven't actually recognized which parts of the image are a car, nor what pieces it's made of. Instead they have memorized some textures and simple shapes which go with the label. It provides the kind of knee-jerk reaction that allows your brain to make you jump when a large object approaches fast, or think there's a tiger hidden in the dirty laundry when you turn around in a dark room. That's why, when you reverse the process, it doesn't create meaningful images, but clumps of relatively common textures found in the training set. It lacks the hierarchy of concepts that allows you to identify objects and distinguish them from the background, which a baby learns in their first two years.