Unpopular opinion time:

A ML model is clearly a derivative work of its input.

Here's what I think would be fair:

Anyone who holds copyright in something used as part of a training corpus is owed a proportional share of the cash flow resulting from use of the resulting models. (Cash flow, not profits, because it's too easy to use accounting tricks to make profits disappear).

In the case of intermediaries (e.g., social media like reddit & twitter) those intermediaries could take a cut before passing it on to the original authors.

Obviously hellishly difficult to administer so it's unlikely to happen but I don't see a better answer.