What is Goodhart's Law, and why do metrics go wrong?

Goodhart's Law is usually stated as: when a measure becomes a target, it ceases to be a good measure. When a measure is used to reward performance, people have an incentive to manipulate it. This can improve the measured score while actually reducing the effectiveness of the system being measured. The common wording is a later compression of the original idea, which was narrower and concerned monetary policy. [1] [2]

British economist Charles Goodhart described the problem in the 1970s while examining monetary growth targets used to guide fiscal policy. His original form read: any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes. The best-known phrasing is generally credited to anthropologist Marilyn Strathern, and the original remark began as something of a jocular aside before spreading widely. [3]

A related finding came from social scientist Donald Campbell, who appears to have identified the concept first and focused on how metrics lead participants to exploit them. Goodhart was instead describing a structural breakdown in inference when rules change, a precursor to the Lucas critique in economics. Keith Hoskin is credited with another paraphrase: every measure which becomes a target becomes a bad measure. [4]

One explanation is that a metric can be accurate before anyone targets it, but once decisions depend on it, it becomes part of the causal system. Employees change workflows, managers select cases, and institutions redesign products. The earlier correlation was measured in a world that no longer exists. Before targeting, a measure tracks the goal with ordinary error; afterward, it also reflects responses and gaming. [2]

The same logic applies to AI evaluation. A Stanford textbook notes that measurement analysis often assumes the act of measuring does not change the phenomenon. Once benchmark scores affect reputation, funding or regulatory standing, developers have reason to optimize for the benchmark rather than the broader capability, and the textbook calls this a demonstrable, recurring phenomenon. [5]

One analysis lists four high-level causes. First, easy-to-measure data is rarely the same as important data; self-reported diet is easy to collect but inaccurate, and the h-index was largely chosen because citation data was available. Second, a metric correlated with a goal becomes less correlated at extreme values: height and basketball skill relate, but the tallest players are not reliably the best. [4]

Third, incentives tied to a metric that does not cause the goal can be perverse. The authors give the example of a teacher grading on the number of questions asked, which may encourage questions that detract from learning. Fourth, the goal itself may be unclear, incoherent or in conflict, as when a committee holds differing values, or when outcomes arrive too slowly to be measured directly. [4]