That makes a lot of assumptions about the nature of the content provided by the recommendation algorithms, as well as human nature. A good recommendation engine, for example, would recognize when someone either likes a broad range of sources politically speaking, or likes a neutral region.

Conversely, it's unclear that a recommendation engine would be able to predict what would be best at "disrupting an echo chamber", and more importantly, when that is desirable, and what "desirable" even means. It's also unclear that the first model is successful at all in disrupting echo chambers, as opposed to exacerbating or amplifying existing positions. I think there's good reasons to think that provocative can be less effective if anything.