Initially this won't check for individual spam messages, but will evaluate accounts. So any comments from 'spammer' accounts would be hidden (or marked). It uses some machine learning approaches which are actually easier to apply than they are to explain - though it's not perfect.