Feedback Loops & Filter Bubbles

The loop is mechanical: the system shows items → users engage with what was shown → the model trains on those engagements → it shows more of the same. Popularity compounds (rich-get-richer), niches starve (nothing shown → nothing learned → never shown).

The filter bubble is the same mechanism viewed from inside a user: the model's belief 'this user only likes X' is unfalsifiable in its own data — it never observes the user reacting to anything else. The bubble isn't a social phenomenon bolted onto a neutral model; it's the default dynamical behavior of a system that trains on its own output.