Goodhart's law is the observation that a statistical relationship may break down when it is used as a basis for control or decision-making. Originally formulated in the context of monetary policy, it is widely expressed as: “When a measure becomes a target, it ceases to be a good measure.” The underlying problem is that an indicator selected because it tracks a desirable outcome may no longer track that outcome reliably once people or systems are rewarded for increasing the indicator itself. Despite its name, Goodhart's law is an empirical generalization and analytical warning, not a universal mathematical theorem.
Origins and meaning
The principle is named after British economist Charles Goodhart, who presented it in 1975 in a discussion of monetary management. His original formulation was:
“Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.”
Its initial setting concerned the use of monetary aggregates as policy targets. A central bank might observe a relatively stable relationship between a measure of the money supply and economic activity or inflation. However, attempts to control that monetary measure can change the behavior of financial institutions and the public, weakening the relationship on which the policy relied.
For example, targeting a particular category of bank deposits can encourage the development or use of financial instruments outside that category. The targeted aggregate may then become less informative about overall monetary conditions. This illustrates how financial innovation and changes in behavior can make an indicator less reliable without requiring anyone to falsify data.
The shorter, widely circulated formulation about a measure becoming a target is associated with anthropologist Marilyn Strathern's 1997 discussion of audit and evaluation. It broadens the principle beyond its original monetary setting to organizational performance, public administration, research assessment, and other systems of measurement.
The two formulations emphasize related but distinct problems. The original concerns the instability of statistical relationships under intervention; the popular version emphasizes the distortion of indicators under target-driven behavior.
How targeting changes a measure
A performance measure usually represents only part of a broader objective. Examination scores do not capture every dimension of learning, publication counts do not fully represent research quality, and recorded output does not necessarily capture the value or durability of production.
Before a measure becomes a target, it may correlate with the desired outcome because ordinary activities improve both. Once rewards or penalties are attached to it, people have reasons to seek ways of improving the measure that may not improve the underlying outcome.
Several mechanisms can produce this divergence:
- Effort substitution: attention shifts toward measured activities and away from unmeasured ones.
- Selection: organizations choose easier clients, cases, or projects to improve reported performance.
- Reclassification: activities are relabeled so that they fall inside a rewarded category or outside a penalized one.
- Reporting manipulation: records are altered, omitted, or timed to improve the indicator.
- Structural adaptation: institutions and markets change in response to the target, making an earlier statistical relationship less applicable.
These mechanisms need not involve dishonesty. Staff may comply exactly with the stated rules while producing an outcome that differs from the rules' intended purpose.
The problem is closely connected to the principal–agent problem. A principal wants an outcome that is difficult to observe, while an agent is rewarded according to an observable proxy. When the agent can influence the proxy independently of the desired outcome, their incentives can diverge. Contract theory studies this difficulty, including how strong incentives for measurable tasks can divert effort from tasks that are important but harder to measure.
A statistical illustration
Goodhart-type effects can arise even without strategic behavior. Suppose a measured score equals a true quantity of interest plus measurement noise:
Assume that and are independent, normally distributed, and have means of zero, with variances and . Then the conditional expected value of the true quantity is
When measurement noise is present, the expected true quantity is less extreme than the observed score. Selecting the highest measured scores therefore selects a mixture of genuine performance and favorable measurement error.
This is related to regression to the mean. It does not imply that selecting on the measure is useless: higher scores still predict higher true values under these assumptions. Rather, treating the measured gain as entirely genuine overstates the improvement. More severe failures can arise when optimization also changes the relationship between the score and the objective.
Main forms
A useful analytical classification distinguishes four forms of Goodhart-type failure. They can occur together and are not a single, universally adopted taxonomy.
Regressional effects
An indicator combines information about the desired outcome with noise. Selection for unusually high indicator values also selects unusually favorable noise, so underlying performance is less exceptional than the measurement suggests.
Extremal effects
A relationship that holds within an ordinary range may fail at extreme values. Aggressive optimization can push a system beyond the conditions under which its indicator was validated.
For example, a production rate may indicate efficiency under normal workloads but cease to do so when further increases generate equipment damage or defective output.
Causal effects
An observed association does not necessarily establish causation. An indicator may be an effect of success, share a common cause with success, or be connected to success through a mechanism that an intervention changes. Directly increasing the indicator may therefore fail to produce the desired result.
Adversarial effects
People or automated systems deliberately exploit a scoring rule. They identify actions that receive high scores while contributing little to the intended objective, or while undermining it.
This is the form most closely associated with “gaming the system,” but Goodhart's law is broader than deliberate gaming.
Applications
Monetary policy and financial regulation
Goodhart's original application concerned the reliability of monetary aggregates as policy targets. More generally, regulation can change the structure of the activities it measures.
In finance, institutions may respond to a regulated risk measure by changing portfolios, legal structures, or accounting treatment. A reported indicator can improve without an equivalent reduction in underlying risk. Whether this occurs depends on the measure, the rules, and the available responses; regulatory ratios do not inevitably fail.
Organizational and public-sector performance
Targets for output, processing speed, sales, or case completion can focus effort and improve accountability. They can also encourage the neglect of quality, maintenance, difficult cases, or longer-term outcomes.
A hypothetical service center rewarded only for short call durations may reduce recorded handling time by transferring difficult cases or ending calls prematurely. The example illustrates a distinction between meeting the indicator and fulfilling the service objective.
Education and research
In education, high-stakes testing can encourage instruction concentrated on tested material rather than the broader curriculum. Higher test scores may reflect genuine learning, familiarity with test formats, curriculum narrowing, or some combination of these effects.
In research assessment, publication counts and citation-based indicators measure aspects of scholarly activity but do not exhaustively represent originality, reliability, or importance. Strong incentives tied to a single metric can influence project selection and publication practices.
These examples do not show that testing or research indicators have no informational value. They show why an indicator's interpretation depends partly on the incentives surrounding its use.
Economic and social indicators
Gross domestic product measures economic production, not every component of social welfare or quality of life. Policies focused exclusively on increasing it may neglect outcomes outside its scope.
This is partly a problem of incomplete measurement rather than necessarily a Goodhart effect. The stronger Goodhart-type claim arises when targeting the indicator changes behavior in ways that weaken its relationship with the broader objective it was intended to represent.
Artificial intelligence
In machine learning, a model is optimized using a loss function, reward signal, or evaluation score that represents intended performance only imperfectly.
A system can achieve a high score by exploiting features of the evaluation rather than performing the intended task. Such behavior is often discussed as reward hacking. Stronger optimization can expose weaknesses in a proxy that were inconspicuous under weaker optimization.
Not every machine-learning failure is a Goodhart effect. Poor generalization, flawed training data, or implementation errors can occur without a proxy being strategically exploited. The relevant connection is the divergence between optimizing a measurable objective and achieving the intended outcome.
Related concepts
Campbell's law concerns how the use of quantitative social indicators for decision-making can create pressures that corrupt both the indicators and the processes they monitor. It overlaps substantially with the broader organizational interpretation of Goodhart's law.
The Lucas critique argues that economic relationships estimated under one policy regime may not remain stable under another because people's expectations and decisions change. It is closely related to Goodhart's original emphasis on the instability of observed relationships under policy intervention, though the two principles have different formulations and scopes.
Goodhart's law also informs mechanism design: the rules of a system must account for how participants respond to incentives, rather than assuming their behavior remains unchanged after a rule is introduced.
Limits and implications for measurement
Goodhart's law does not establish that every target becomes useless, that all performance measurement is harmful, or that incentives necessarily worsen outcomes. A measure can remain informative after becoming a target, especially when it is closely connected to the objective and difficult to manipulate independently of it.
Its practical significance depends on:
- the gap between the indicator and the underlying objective;
- the strength of the incentives attached to the indicator;
- the scope for changing behavior without improving the outcome;
- the stability of the environment in which the indicator is used;
- the presence of independent checks on quality and unintended effects.
Common institutional responses include using several complementary measures, distinguishing monitoring indicators from reward targets, checking reported performance against independent evidence, and revising metrics as behavior changes. These approaches involve trade-offs: additional measures can increase complexity, while discretionary assessment can introduce inconsistency or bias. No fixed collection of indicators eliminates the need to examine whether measured improvement still represents the outcome being sought.