All of these models retain a representation of the original training data in their parameters, which makes training a violation of copyright unless it was explicitly authorized

I don’t think it is realistic to train a AI model by specific authorised database.