I've thought about this as well. Privacy is actually one impediment I think. If you make everyone's raw usage data public (so that anyone can use it to train their own recommendation model), some of the training data might be de-anonymizable.

However there's plenty of data out there that people are already happy to make public, so the privacy thing probably isn't a show-stopper. e.g. Reddit seems like it'd be a rich source of training data. Twitter too, if it were more open. I'd keep an eye on Bluesky!