The only way I see this working with out current economics and IP law is if the people training models license the work they are using. And a traditional license wouldn't do, it would have to be one specific to training AI models.

As to the question of worth, obviously OpenAI's models have value without the training data. Just having a collection of images does not make a trained AI. But the total value of the system is a combination of that model and the training data.