I think what it comes down to is that most of the other ways to automate ID creation at scale were probably really easy to detect using ... well, probably neural networks. Because ID verification services had access to millions of "real" ID photos, I'd think a lot of fake images created using traditional automation would be pretty easy to detect (For example, having the same backgrounds/wood grain pattern reused for multiple images). So what neural network generation is adding here is basically being able to add much much more "plausible randomness" to the image (lens artifacts, dust motes, wood grain, plastic reflection, etc) in a way that makes distinguishing generated from non-generated images statistically very difficult.