It'd be dead easy, actually. Don't even have to actually ban it: For image generating models, every artist whose work is included in the training data becomes entitled to 5 cents per image in the training data every time a model generates an image, so an artist with 20 works in the model is entitled to a dollar per generated image. Companies offering image generating neural networks would near instantly incur such huge liabilities that it simply wouldn't be worth it anymore. Same thing could apply to text and voice generating models, just per word instead of per image.