If you never get to see what the computer thinks you are not going to like, as soon as its model of you becomes slightly wrong, you will never have the chance of correcting it and getting to see the good stuff.
This is an interesting technical objection. I am not sure that it is necessarily true, though. If I properly weigh exploration (vs. exploitation), and I present related users with stochastic feeds, then the system can explore things it didn't realize that I might like.
Not that I think genetic algorithms are useful in general, but a genetic algorithm would be useful in this specific circumstance. If I engage in K discussions when the feed is driven by exploitation, but I engage in 2K discussions when the feed is driven by more exploration, then I will prefer explorative feeds. So I believe it would be possible to escape local minima.