The only thing that matters in this case is the types of devices used to make the pictures in the training data, and the device being used to make the picture t

The only thing that matters in this case is the types of devices used to make the pictures in the training data, and the device being used to make the picture that is analyzed by the algorithm. Advances in computational photography software in “flag ship phones” is not really relevant. Also both sides of this argument are making the mistake of equating photos that are visibly recognizable to a human, with a photo what would provide enough distinct features that an algorithm could make an ID prediction. There could be some bias in when people assembled the data set they chose humanly recognizable images, but it’s not necessary. Also it’s obvious that the kind of features on a pixel level that would be identifiable on a photo of a dark complexion and a fair one would be different. so either you train separate algorithms or you make sure your training data set has enough examples of both.