I don’t understand — what’s the control variable here? How can they claim the algorithm has a deradicalizing effect without showing that people became less radi

I don’t understand — what’s the control variable here? How can they claim the algorithm has a deradicalizing effect without showing that people became less radical? Also, the raw number of recommendations is misleading — as long as 1 in 100 of the recommendations is a new piece of content that is also radical, I could be exposed to radical beliefs over time just by my own selection bias.