I wonder how much of this is due to the fact that initially the system really is mostly trained by your preference, but over time, it then becomes trained on yo

I wonder how much of this is due to the fact that initially the system really is mostly trained by your preference, but over time, it then becomes trained on your preference mediated by the algorithm, and then slowly you see a divergence of your actual preferences and what it suggests to you/you end up clicking on. Peter Norvig had a great lecture about this in the context of language translation, and how despite so much info being out there on the web, known "good" corpuses of data that were 100% human-made and guaranteed to not being the product of an algortihm (like official translations of UN documents) were still super useful.