Glass needs to operate on data in-flight, what better way to train it? However, you do have a point wrt to Youtube. I think I left my <cynicism> tag out of the previous post.
I don't know why you are getting down-voted. It is in Google's interest to have a large training corpus of faces for many machine learning algorithms (face recognition, expression recognition, age detection, etc.).
Training data doesn't need to be in-flight. Testing/operational data does. These are two different things, and the GP makes a good point.