<- Back to laws

Measurement and control principle

Goodhart's Law

A measure can describe a system well and still become unreliable when rewards, penalties, or decisions are tied to it. The target changes behavior; behavior changes the relationship the measure was meant to capture.

TypeSocial-science heuristic
DomainMetrics & incentives
Scientific statusGeneral pattern, not theorem
Core mechanismBehavioral response
Evidence baseEconomic & institutional
Common misuseRejecting all measurement
INTERACTIVE 01 / METRIC DESIGN LAB

Watch a useful measure become a dangerous target.

This teaching model separates reported score from true outcome. Target pressure creates incentives to improve both real performance and the appearance of performance. Audits and counter-metrics make gaming costlier and harder to hide.

Illustrative, not predictiveThe curves explain a mechanism. They do not estimate any real organization.
ObservationPay / punishment
WeakStrong
One numberBalanced set
LIVE SYSTEM MODELHIGH DISTORTION RISK
Reported score83what the dashboard sees
True outcome50what the mission needs
Gaming share39%effort spent on the proxy gap
Interactive visual model for Goodhart's Law.
Reported scoreTrue outcomeGaming share

At this pressure, the dashboard is improving much faster than the underlying outcome. The metric is now partly measuring adaptation to the target.

01ObserveA measure correlates with a valued outcome.
02TargetRewards or sanctions attach to the measure.
03AdaptPeople optimize what the system can see.
04DivergeThe score and the purpose separate.
01 / MEANING

The law is about intervention, not bad statistics.

The familiar formulation is: "When a measure becomes a target, it ceases to be a good measure." That sentence is a later, memorable compression. Charles Goodhart's original monetary-policy observation was narrower: statistical regularities used for control tend to break down under the pressure of control.[1][2]

The distinction matters. A metric can be accurate before it is targeted. Once decisions depend on it, the metric enters the causal system. Employees change workflows, customers respond to rules, managers select cases, and institutions redesign products. The old correlation was measured in a world that no longer exists.

BEFORE TARGETINGM = G + noiseThe measure M tracks the goal G with ordinary error.
->
AFTER TARGETINGM = G + response + gamingThe policy creates new causes of the measured value.
MeasureA visible proxy: test score, sales count, wait time, clicks.
GoalThe underlying purpose: learning, service, health, trust.
Control pressureRewards, sanctions, ranking, funding, or reputation.
Key diagnosticAsk whether the act of using the metric changes the process that generated it. If yes, historical validity may not survive policy use.
"A dashboard is not outside the system. Once people can see the target, the dashboard becomes part of the machinery."
02 / ORIGIN & EVOLUTION

From monetary aggregates to audit culture.

Goodhart developed the idea in the context of monetary management, where relationships involving a chosen monetary aggregate could become unstable once authorities attempted to control that aggregate. The principle later traveled into public administration, education, corporate management, and machine learning.

1975Policy control

Goodhart presents the core observation in work on U.K. monetary management at a Reserve Bank of Australia conference.[1]

1979Corruption pressure

Donald Campbell describes how quantitative indicators used for social decisions become vulnerable to corruption and distortion.[4]

1997Audit culture

Marilyn Strathern popularizes the concise target formulation while examining accountability and evaluation.[3]

2018Mechanism taxonomy

Manheim and Garrabrant distinguish regressional, extremal, causal, and adversarial forms.[5]

Wording noteThe popular sentence is useful, but it should not be presented as a verbatim 1975 quotation. It is a later formulation of the underlying idea.
03 / FOUR FAILURE MECHANISMS

There is more than one way for a proxy to fail.

The modern taxonomy below is analytical, not a claim that every case fits one box. Real systems often combine several mechanisms.[5]

01 / REGRESSIONAL

Selecting the lucky tail

The proxy combines the goal with noise. Selecting extreme proxy values also selects unusually favorable noise, so expected goal quality is lower than the score suggests.

Signal: top-ranked cases regress on retest.
02 / EXTREMAL

Leaving the known regime

A relationship holds inside the observed range. Optimization pushes the system into unfamiliar conditions where omitted variables or nonlinear effects dominate.

Signal: the model is extrapolating.
03 / CAUSAL

Moving the proxy, not the goal

The measure predicts the outcome because both share a cause, or because the goal causes the measure. Directly manipulating the measure does not reproduce the desired causal pathway.

Cause->Measure / Goal
Signal: intervention breaks correlation.
04 / ADVERSARIAL

Optimizing against the evaluator

An agent understands the rule and changes visible behavior to score well while preserving a different private objective.

RULEVERSUSRESPONSE
Signal: performance drops when audit changes.
04 / DIAGNOSTIC MAP

Pressure and proxy quality determine the risk zone.

A weak proxy is not automatically useless, and a strong proxy is not automatically safe. The danger increases when narrow measures carry high consequences and affected people can adapt faster than the measurement system.

01Can people move the score without moving the outcome?
02What valuable work is invisible to the metric?
03Who bears harm when the target is met?
04Would the relationship survive a policy intervention?
05How quickly can behavior adapt to the rule?
06What evidence could reveal metric decay?
05 / DOCUMENTED CASES

When the number displaced the purpose.

These cases illustrate incentive and measurement failure; they do not imply that Goodhart's Law alone explains every organizational cause.

MONETARY POLICY / 1970s

Controlling the indicator changed the relationship.

GoalStable monetary conditions->ProxyChosen monetary aggregate->ResponseFinancial behavior adapts->FailurePrior regularity weakens

This is the law's home territory. A monetary relationship observed under one policy regime cannot simply be assumed invariant after authorities make it an instrument of control.[1][2]

MechanismsCausal + extremal
RETAIL BANKING / 2016

Sales goals rewarded accounts, not customer value.

GoalUseful customer relationships->ProxyProducts sold->ResponseAccounts opened without consent->FailureScore rises, trust collapses

The CFPB found that sales targets and compensation incentives spurred Wells Fargo employees to open unauthorized accounts. A later independent investigation described a sales culture and control failures surrounding the practices.[6][7]

MechanismsAdversarial + causal
EDUCATION / ACCOUNTABILITY

Test security became part of system design.

GoalStudent learning->ProxyStandardized test results->ResponseTeaching, exclusion, or cheating->FailureScore can overstate learning

Campbell warned that high-stakes social indicators face corruption pressure. GAO later documented the importance and variation of state test-security policies, audits, and statistical monitoring in accountability systems.[4][8]

MechanismsAdversarial + regressional
CALL CENTERS

Short calls, unresolved problems

Average handling time improves while repeat contacts and customer effort increase.

PLATFORMS

Engagement without welfare

Clicks and time-on-site rise even when users feel worse about the experience.

SOFTWARE

Velocity without value

Ticket counts rise as work is split, low-value tasks are favored, or quality debt is deferred.

06 / METRIC GOVERNANCE

Do not abandon metrics. Design their use.

The practical response is a measurement system that treats validity as something to maintain, not a property established once. Metrics should inform judgment, expose tradeoffs, and remain open to challenge.

01Define purpose

Write the real outcome before choosing a number.

02Map the proxy

State why the metric should track that outcome.

03Model response

Predict how each affected group can adapt.

04Add counter-metrics

Measure quality, harm, equity, and long-term effects.

05Audit independently

Compare reported scores with samples and outcomes.

06Revalidate

Retire or revise measures when behavior changes.

PRIMARY METRICWhat progress looks like
+
COUNTER-METRICWhat must not be sacrificed
+
GUARDRAILWhat is never acceptable
+
REVIEW TRIGGERWhen validity is tested again
Strong rule of thumbThe higher the stakes, the less defensible it is to let one visible number make the whole decision.
07 / LIMITS & MISUSE

Goodhart's Law is a warning, not a veto.

Not every target destroys its measure. Targets can coordinate action and improve performance when the metric is causally connected to the outcome, hard to manipulate, balanced by other evidence, and reviewed as conditions change.

Misuse: metric nihilism

"All numbers become useless."

Better: reduce stakes, triangulate, audit, and revalidate.
Misuse: automatic accusation

"A rising score proves gaming."

Better: look for independent outcome evidence and response mechanisms.
Misuse: clever slogan

"Goodhart explains every KPI failure."

Better: identify the specific causal, statistical, or adversarial mechanism.
Misuse: no accountability

"Professional judgment needs no measurement."

Better: combine accountable judgment with plural evidence.
09 / REFERENCES

Sources and further reading.

Primary publications, official archives, regulatory records, and peer-reviewed work are prioritized.

  1. Reserve Bank of Australia - Papers in Monetary Economics bibliographyOfficial bibliographic record for Goodhart's 1975 conference papers, including Problems of Monetary Management: The U.K. Experience.rba.gov.au/publications/rdp/1990/9013/conference-volumes.html
  2. Bank of England Archive - Monetary policy papers, 1975Archival context for monetary-target analysis and Goodhart's policy work.bankofengland.co.uk/CalmView/Record.aspx?id=6A50%2F17
  3. Marilyn Strathern - Improving Ratings: Audit in the British University SystemThe 1997 discussion associated with the now-familiar target formulation.doi.org/10.1111/j.1467-954X.1997.tb03453.x
  4. Donald T. Campbell - Assessing the Impact of Planned Social ChangeFoundational treatment of corruption pressure and distortion in high-stakes social indicators.doi.org/10.1016/0149-7189(79)90048-X
  5. David Manheim & Scott Garrabrant - Categorizing Variants of Goodhart's LawFormal taxonomy of regressional, extremal, causal, and adversarial mechanisms.arxiv.org/abs/1803.04585
  6. Consumer Financial Protection Bureau - Wells Fargo enforcement actionOfficial 2016 account of unauthorized accounts linked to sales targets and compensation incentives.consumerfinance.gov/archive/newsroom/...wells-fargo...
  7. Wells Fargo Independent Directors - Sales Practices Investigation ReportInvestigation findings filed with the SEC on sales culture, management, and control-function failures.sec.gov/Archives/edgar/data/72971/.../d375947ddefa14a.htm
  8. U.S. GAO - States' Test Security Policies and Procedures VariedOfficial review of testing security, audits, statistical analysis, and detected cheating in accountability systems.gao.gov/products/gao-13-495r
  9. Hennessy & Goodhart - Goodhart's Law and Machine Learning: A Structural PerspectiveA modern economic model of prediction, manipulation costs, and Goodhart bias.doi.org/10.1111/iere.12633
CONTINUE EXPLORING

Related laws, with the relationship made explicit.

These are editorial connections, not claims that the laws are mathematically equivalent.

CONTINUE READING

Place this law inside the collection.

LAW 003 / 100 PUBLISHED