I imagine this is a corollary of the phrase, "you manage what you measure." Usually the metrics we measure are an indirect substitute for what we actually want

I imagine this is a corollary of the phrase, "you manage what you measure." Usually the metrics we measure are an indirect substitute for what we actually want to measure, e.g., lines of code as a measure of programmer productivity. What inevitably happens is that behaviour is optimised for the metric, and not what the metric was intending to measure. It's really just gaming the system.