I tested those "cloud face" in the popular face recognition engine rekognition.com. None of them actually got recognized as face. Looks like the face recognition algorithm they uses are pretty smart.
If you upload the image of the overall piece (the one with an out-of-focus head looking towards it) 2 of the cloud faces are recognized as faces. The other image of the piece taken at a more skewed angle with no head blocking it actually recognizes 3 of the panels as faces.
The funny thing about seeing false faces (or patterns in general, really) is that scale (and to some degree composition) matters, even for us humans. When I look at the large picture of the cloud that is the first one presented in the article at 100% scaling, I struggle to see any sort of meaningful face. But when I locate that same image as a panel on the overall work in a smaller scaled down presentation I see the 'face' immediately. With a browser that can appropriately scale down to 25% or less I can see the same thing happening with the actual large source images... once I hit about 50% scaling I can see the "faces" that the computer recognized.
Obviously YMMV depending upon monitor size vs dpi as to where they stop being clouds and start being faces.
Different algorithms will usually get different false positives. This is why large collections of algorithms generally perform much better than the best individual algorithm. I imagine these are all weird edge cases as it is.
