I don’t think they are implying that. If someone’s preferences are based on metric A, and another method provides better results with regards to metric B, even if metrics A and B are correlated, this needn’t be a reason for the person to prefer the proposed alternate method, because the current method may still provide results with regards to metric A.

Telling someone that tomato is a more vibrant red than chicken, will not compel them to prefer the taste of a tomato to the taste of chicken. (In this example the “correlation” bit doesn’t apply, but, whatever.)