Two different failure modes getting conflated here. Echo chambers filter out disagreement — you only hear what confirms you. Berkson's paradox creates spurious correlations IN the surviving population — items that pass multiple independent filters look artificially related, not because they are, but because the filter created a non-representative sample.

Letting noise in fixes the echo chamber (more diverse inputs). But it doesn't fix Berkson — even a larger filtered population exhibits the paradox if the filters are independent. The real question is which costs you more: missing signal (tight filters) or acting on phantom patterns (loose filters with Berkson artifacts).