a single person produces enough data in my mind to train an AI for him or her. We just need to apply existing algorithms to the problem. We just need a lot of data to make a generic AI that works for all. Which isnt the case for the above proposal
The problem isn't the algorithms, the problem is the labeled training sets which those algorithms need to create models.
Also, to recommend news stories a la Google Now you need to scrape an appreciable portion of the Internet and extract features to score against your model.