Maybe what Schmidt said about Google's intentions was true, but another reason they haven't done it is because it's not possible right now, or at least not in a fully-automated way across most images on the web. The state-of-the-art on the easier "verification" problem ("are these two images of the same person?") are shown here:
http://vis-www.cs.umass.edu/lfw/results.html
The best results are under 90% accuracy, which sounds pretty good, until you realize that random chance is 50%, and for recognition ("who is this person?"), you're essentially exponentiating that 90% by the number of different people you want to recognize.
