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.
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.
At this pressure, the dashboard is improving much faster than the underlying outcome. The metric is now partly measuring adaptation to the target.
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.
"A dashboard is not outside the system. Once people can see the target, the dashboard becomes part of the machinery."
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.
Goodhart presents the core observation in work on U.K. monetary management at a Reserve Bank of Australia conference.[1]
Donald Campbell describes how quantitative indicators used for social decisions become vulnerable to corruption and distortion.[4]
Marilyn Strathern popularizes the concise target formulation while examining accountability and evaluation.[3]
Manheim and Garrabrant distinguish regressional, extremal, causal, and adversarial forms.[5]
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]
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.
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.
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.
Optimizing against the evaluator
An agent understands the rule and changes visible behavior to score well while preserving a different private objective.
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.
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.
Controlling the indicator changed the relationship.
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]
Sales goals rewarded accounts, not customer value.
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]
Test security became part of system design.
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]
Short calls, unresolved problems
Average handling time improves while repeat contacts and customer effort increase.
Engagement without welfare
Clicks and time-on-site rise even when users feel worse about the experience.
Velocity without value
Ticket counts rise as work is split, low-value tasks are favored, or quality debt is deferred.
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.
Write the real outcome before choosing a number.
State why the metric should track that outcome.
Predict how each affected group can adapt.
Measure quality, harm, equity, and long-term effects.
Compare reported scores with samples and outcomes.
Retire or revise measures when behavior changes.
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.
"All numbers become useless."
Better: reduce stakes, triangulate, audit, and revalidate."A rising score proves gaming."
Better: look for independent outcome evidence and response mechanisms."Goodhart explains every KPI failure."
Better: identify the specific causal, statistical, or adversarial mechanism."Professional judgment needs no measurement."
Better: combine accountable judgment with plural evidence.Sources and further reading.
Primary publications, official archives, regulatory records, and peer-reviewed work are prioritized.
- 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
- 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
- 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
- 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
- David Manheim & Scott Garrabrant - Categorizing Variants of Goodhart's LawFormal taxonomy of regressional, extremal, causal, and adversarial mechanisms.arxiv.org/abs/1803.04585
- 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...
- 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
- 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
- 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