Charles Goodhart’s 1975 original concerned monetary policy (“any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes”); the famous punchy phrasing, “when a measure becomes a target, it ceases to be a good measure”, is Marilyn Strathern’s 1997 generalisation. Either way, the mechanism is the same: a metric is a proxy for something you actually want, and applying optimisation pressure finds the cheapest way to move the proxy, which is rarely the intended way.
In machine learning the law bites twice: models exploit loopholes in their loss functions, and teams exploit loopholes in their dashboards. The defence is the same at both levels: red-team the metric by asking what degenerate behaviour scores well, patch what you can into the metric itself, and surround the rest with guardrail metrics that catch the exploits the headline number invites.