Goodhart's Law

Descriptive

Full Practice · Knowledge · Reading What's Operating

Mechanism

Goodhart's Law reads the moment a measure stops being a window and becomes a handle.

A metric begins as a proxy for something you care about. Test scores stand in for learning. Response time stands in for service quality. Citations stand in for intellectual contribution. Benchmark performance stands in for model capability. The measure is never the thing itself, but while pressure is low it may track the thing well enough to help you see.

Then the system starts optimizing the measure. Funding, status, promotion, punishment, publication, release gates, or legitimacy attach to the number. People adapt. They teach to the test, close tickets without fixing the problem, slice research into publishable units, tune the model to the benchmark, redefine the category, delay reporting, or learn exactly where the audit will look.

Goodhart's Law reads the pressure that breaks the relationship between proxy and reality.

The metric did not become useless because measurement is bad. It became unreliable because the system began acting on the proxy as if it were the target. Once the number controls consequences, the number becomes part of the system it was supposed to describe.

Goodhart's Law is close to Mechanism Design because every target creates a game. It is close to Report Fidelity because reports fail when the object of reporting becomes detached from what the report is used to claim. It is close to Legibility because institutions often choose measures because they are administratively visible, not because they are the best contact with reality.

Control misreads Goodhart's Law by adding more targets, tighter dashboards, harsher audits, and more elaborate compliance machinery. The system responds by learning the new numbers. Decay misreads it by treating every metric as corrupt and abandoning measurement altogether. The Range reading keeps measurement answerable to the thing it claims to measure, and keeps the consequences attached to the number modest enough that the number can still tell the truth.

Practice

The diagnostic question is: "What behavior does this metric reward once people know they are being judged by it?"

Use this when a number, benchmark, checklist, score, ranking, audit, KPI, or threshold begins to shape behavior rather than merely describe it.

The practice is not anti-measurement. Bad measurement is not fixed by pretending you can see without instruments. The fix is to keep the instrument in its place. A measure should discipline attention, not replace judgment.

In the Wild

A school wants better learning. It ties teacher evaluation to test scores. Teachers respond by narrowing instruction around the test, avoiding difficult students, and training test-taking behaviors that raise the score without deepening understanding. The number may rise while the thing it stood for weakens.

A support team is measured by ticket closure time. Agents learn to close quickly, split hard cases into smaller tickets, or move unresolved issues into categories that do not count against them. The dashboard improves. The customer experience does not. The metric has become an incentive for making the problem less visible.

An AI lab tracks benchmark performance. The benchmark begins as a useful external test. Then it becomes a release target, a marketing claim, and a competitive signal. Models are tuned against it, prompts are optimized for it, and capability begins to look better on paper than in open conditions. The benchmark still tells you something. It no longer tells you what it told you before pressure attached to it.

Goodhart's Law asks you to respect the distance between the number and the world. Count what helps you see. Then keep asking what the counting is teaching people to do.

Lineage

The Codex did not invent Goodhart's Law. It inherits the tool from economics, policy evaluation, social science, and later AI alignment work.

Charles Goodhart formulated the core idea in monetary-policy terms while analyzing the United Kingdom's experience with monetary aggregates: statistical regularities used for control tend to break down under control pressure. The original context matters. The problem was not a generic complaint about numbers. It was a warning about governing through observed relationships after policy pressure has made those relationships unstable.

Donald Campbell articulated a closely related law about social indicators: the more a quantitative indicator is used for social decision-making, the more it is subject to corruption pressure and the more it can distort the process it was meant to monitor. Marilyn Strathern carried the measure-becomes-target formulation into audit-culture critique, citing the accountability lineage around Goodhart's law rather than originating the monetary-policy warning herself.

The surrounding lineage includes the Lucas critique in macroeconomics, principal-agent theory, audit studies, performance management, psychometrics, and public administration. All point at the same family of failures: measurement changes behavior when consequences attach to measurement.

AI alignment and machine-learning researchers have made the tool sharper for model evaluation. Benchmark overfitting, reward hacking, proxy objectives, specification gaming, and distribution shift are Goodhart-family problems. The measure may correlate with the intended goal in ordinary conditions and then fail under optimization pressure.

The Codex uses Goodhart's Law here as a reading instrument. Before changing the dashboard, incentive, or rule, you need to see the metric-pressure field: what the institution is making legible, what actors are optimizing, and where the proxy has drifted from the thing.

The tool has limits. Goodhart language can become lazy anti-metric rhetoric. A metric being imperfect does not make it useless. A target being gameable does not mean the system should return to unmeasured judgment. The better question is whether the metric still helps you see once pressure is attached, and what other contact with reality keeps it honest.

Cross-references

Within the category. Mechanism Design asks what game the metric creates. Rules-in-Use asks whether the formal target or the adapted behavior actually governs the system. Legibility reads why the metric may have been chosen because it was visible to authority, not because it was faithful to reality.

Across the Workshop. Report Fidelity asks whether a report still supports the claim being built from it. Checking Your Map Against Reality is the natural partner when the proxy begins to replace the thing. Chilling Effects becomes relevant when metrics change speech and behavior before any formal punishment is applied.

Limitation. Goodhart's Law does not say "never measure." It says that measurement under pressure becomes part of the system and must be read as such.