In addition to the caveats they list, there may be systematic biases as to who includes a profile pic and who doesn't. Even an enormous sample and perfect inferences from the face API can't surmount a bad dataset (bad depending on what you are trying to learn from it)
To be fair, they did address that: Lang FacesDetected ruby 71 r 38 javascript 60 java 47 html 59 go 53 cpp 34 c 24 python 49 php 66 perl 45 swift 49 csharp 61
Not quite. That relates to sample sizes and, as far as interpretation goes, it's dealt with as if the rest were missing at random. In reality, they are probably missing not at random, but in relation to other characteristics, changing the interpretation.