Comment by martingalex2

from the article "A pattern we see in some interpretability and interpretability-adjacent ML papers is defining some metric which is claimed to correspond to some property of interest, and then very rigorously measuring this metric. We see this as a kind of Cargo-Cult Science."