The courts will have to settle that according to circumstances. I think it is likely to be the prompter, and in some cases the creator of the training data as well. The creator of the model will have copyright on the model, but unlikely to have copyright on its outputs (any more than the writer of a compiler has copyright on its output).