1) There are many possible implementations. Right now the problem is that very few software engineers understand blockchain sufficiently so that they can design efficient solutions. There is always an overhead when coupling a service to a blockchain but it can be minimal if done correctly.

2) People get paid to provide quality training data... So yeah it's valuable but if there is a built-in financial incentive to share data with the system then people will do it. Also, other people could get paid to verify/curate the training data. The system just needs to setup the right incentives.