Goodhart\'s law

Goodhart\'s law is the observation that when any measure is adopted as a target for policy or optimization, it loses its value as a measure. First articulated by economist Charles Goodhart in 1975 in the context of monetary policy, the law has since generalized across domains as a fundamental principle of systems behavior: the act of optimizing for a proxy severs the correlation between that proxy and the underlying variable it was meant to represent.

The canonical formulation is deceptively simple. Goodhart originally noted that any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes. This is not merely a problem of human perversity or gaming behavior. It is a structural property of feedback systems in which the measured variable is causally coupled to the system being measured. When the system responds to the measure, the measure no longer observes the system from outside; it becomes an endogenous variable, and the statistical relationship that made it useful dissolves.

Mechanisms of Corruption

Goodhart\'s law operates through several distinct but interacting mechanisms:

Gaming and substitution. When teachers are evaluated by test scores, they teach to the test. When police are evaluated by arrest counts, they arrest for minor offenses. When universities are ranked by citation metrics, they encourage citation cartels. The behavior being measured is substituted by behavior that optimizes the measure. This is the surface-level manifestation, and it is the most widely understood.

System adaptation. The deeper mechanism is that the system itself adapts to the metric. In machine learning, this manifests as \'\'reward hacking\'\': an agent discovers an unforeseen strategy that maximizes the reward signal without achieving the intended goal. In economics, it is the \'\'Lucas critique\'\': policy based on historical macroeconomic relationships fails because the policy itself changes the expectations and behavior that produced those relationships. The system is not merely being gamed from the outside; its internal dynamics are being restructured by the presence of the target.

Representational drift. Over time, the gap between the metric and the underlying reality widens not through any single dramatic failure but through accumulated small adaptations. This is the \'\'representational debt\'\' problem: the model (the metric) was once a reasonable simplification of the territory, but the territory has changed in response to the model being used to guide action. The metric becomes a \'\'legacy system\'\' within the organization\'s epistemic architecture — still trusted, still reported, no longer connected to the reality it purports to describe.

The Epistemic Dimension

Goodhart\'s law is not merely a policy problem. It is an \'\'epistemological\'\' problem. It reveals that measurement is never passive observation. Every measurement that guides action becomes an intervention. And interventions, in complex systems, produce side effects that feedback upon the measurement itself.

This creates a paradox for epistemic engineering and resilience metrics. We need metrics to design better epistemic systems. But the act of institutionalizing those metrics risks corrupting the very systems we seek to improve. A metric of \'\'epistemic latency\'\' — the time between the emergence of disconfirming evidence and its integration into organizational belief — seems unambiguously good to minimize. But when organizations are evaluated on this metric, they optimize for speed of belief revision rather than accuracy, producing rapid but superficial updates that look like responsiveness but function as \'\'epistemic theater\'\'.

The problem intensifies in information networks where metrics propagate. A single corrupted metric at one node can cascade through validation channels, not because the network lacks redundancy but because the nodes share the same metric — the same representational debt — creating correlated failure across structurally diverse paths.

Beyond the Law

Responses to Goodhart\'s law fall into two categories, neither fully satisfactory:

The sophisticate\'s response is to build better metrics — more robust, harder to game, more tightly coupled to the underlying goal. This is the approach of \'\'mechanism design\'\' and \'\' Campbell\'s law\'\' research in evaluation science. The hope is that with sufficient cleverness, one can design a metric that survives its own use as a target. This hope is not obviously false, but it has not been obviously fulfilled either.

The skeptic\'s response is to minimize the use of metrics as targets — to rely on judgment, narrative, and qualitative assessment. But this merely relocates the problem. Judgment itself is subject to Goodhart dynamics: when decision-makers know their judgments will be evaluated, they optimize for evaluability rather than accuracy. The \'\'quantified self paradox\'\' extends this to personal life: tracking sleep, mood, or productivity changes the behavior being tracked, often for the worse.

Goodhart\'s law is not a problem to be solved. It is a boundary condition on the possibility of control. Any system complex enough to be worth measuring is complex enough to adapt to being measured. The dream of a metric that cannot be gamed is the dream of a system that cannot learn — and a system that cannot learn is already dead. The question is not how to prevent Goodhart corruption but how to live with it: how to build institutions that expect their metrics to decay and have procedures for retiring them before they become dangerous.