The failure to create human-level AI has been a great disappointment for me.
But it looks like we are a little bit there.
Artificial intelligence has had two main approaches: top-down, by explicit specification, and bottom-up, by learning. The top-down approach seemed to be the way to go at first, but the bottom-up approach has increased in popularity over the last few decades because of ever-increasing computing power.
What Google is doing with those images is essentially bottom-up learning. Those surreal dog-face ones were likely made with a module trained on recognizing dog faces, so that module tends to see a dog face wherever it looks.