I think because this kind of ML is so new, we have no choice but to frame arguments for/against in terms of the structures that have been in place for decades past (copyright, open source licenses). We don't yet have the legal language to express dissent against ML in clear yes or no terms.

I think if there were an option to add a machine learning clause and ask individual creators if they wanted it applied in that context, we would see a considerable amount of uptake. It's just that we couldn't forsee this progress happening so soon, and the issue is still not visible enough. I think it's only a matter of time before the culture catches up and new creative works in the coming years are excluded from training sets by their authors with clear and direct language.

By that point there would be no way to argue "but they shouldn't care, they licensed it like this, so I'm assuming it's fine for ML use."

If copyright is not enough to stop another entity from using a person's data for training, then some other protection should be invented that does.