One method of generating things via neural networks is called a generative adversarial network. It works by having two models. One that generates content and one that detects fake content. You train them both in parallel. As the fake detector gets better, so does the generative model at generating fakes. It’s literally a cat-and-mouse game. If someone came up with a scheme to reliably detect your fakes, you could add it to your discriminator model and retrain the generator to improve the fake generation.