The issue is that the goal recommendation systems in commercial sites is [generally] profits. So, Netflix will not risk to recommend a film outside some confidence level.
In my experience, closer [taste] friends works better than recommendation systems, since they can see outside the typical statistics.
Sometimes it's actually irrational, though. On Pandora, for instance, I find my stations keep playing a relatively small set of songs over and over again (I believe a lot of other people have this problem). From a profit-motive point of view a better algorithm would be exploring the edges of my personally indicated taste space a bit more often but their algorithm is apparently a bit too dumb for this.
The bigger problem, I think, is not insularity caused by algorithms but insularity caused by personal choice. People with political ideology X don't miss out on reading stuff by people with political ideology Y because google doesn't suggest it for 'em, they miss out on reading stuff by people with political ideology Y because they don't want to read it. People live in information bubbles created by themselves, not by matching algorithms.
And the problem is not just that people refuse to step outside their own comfort zone. If I have ideology X and start reading political blogs by the opposing ideology, then I'll quickly get annoyed, because I'll find that these blogs largely consist not of "Here are some sensible arguments in favour of ideology Y" but rather of "hurr durrr people with ideology X are stoooopid and evil! Hey, did you know that people with ideology X say they believe X but in fact they really believe P, Q and R? What a bunch of disgusting freaks!" And so forth.
Usually try to give it 4-5 "goes" at the novel activity to determine if they're good at picking out things that at first aren't interesting.