The dataset used to train the model needs to be sourced ethically, just like the supply chain used by a physical manufacturer needs to be audited to ensure a supplier isn't using slave labour in a country too remote to attract much attention over the issue. In this case, I'd say the companies need to either dilute their datasets further, using fewer samples from any given person to the point that AI can't replicate the appearance of a specific person or the style of an artist except by improbable coincidence or extreme genericity, or get consent from each person who (or whose work) appears in the training data.

Though this is deepfakes, which I think involve users applying additional training material specifically of the target, so that the AI over-fits to that specific output. If the original AI was ethically/respectfully produced, then the people responsible for the additional rounds of training ought to be the ones at fault, at least as much as the prompt-writer themselves (assuming they're not the same individual!). For that, the only good solution I can think of is legislation.