Imagine a factory in the Soviet Union tasked with producing nails and given a quota measured by number. The factory produces millions of tiny, useless nails. The quota is changed to weight. The factory produces a small number of enormous, useless nails. In both cases the metric was met. In both cases the actual goal, useful nails in useful quantities, was entirely missed. The story, which circulated widely in discussions of Soviet central planning, is probably apocryphal. But as an illustration of a principle that shows up in organizations of every kind, it is nearly perfect.
The principle is called Goodhart’s Law, and it is one of those ideas that, once encountered, becomes very difficult to stop noticing. It appears in hospitals, in schools, in tech companies, in governments, and in any environment where people are held accountable to metrics. Understanding it does not make the problem go away. But it does make you considerably harder to fool, including by your own measurement systems.
Contents
The Law and Where It Comes From
Goodhart’s Law is named after Charles Goodhart, a British economist and advisor to the Bank of England, who articulated the principle in a 1975 paper on monetary policy. His original formulation was technical and specific to economic indicators: any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes. The idea has since been restated in more general terms by the social researcher Marilyn Strathern, whose version is the one most widely quoted: when a measure becomes a target, it ceases to be a good measure.
The mechanism is straightforward once you see it. A measure is useful when it is a reliable proxy for something you actually care about. Test scores, as a measure, are useful because they tend to correlate with genuine learning. Hospital wait times, as a measure, are useful because they correlate with quality of patient care. Revenue figures are useful because they tend to correlate with business health. The moment you attach a target to the measure, however, you give people an incentive to optimize for the measure itself rather than the underlying thing it was designed to represent. And optimizing for a proxy, rather than the real thing, has a way of decoupling the two.
The Teaching to the Test Problem
Education offers some of the most familiar examples. Standardized test scores were introduced as a useful proxy for student learning. When school funding, teacher evaluations, and school rankings were tied to those scores, they became targets. At that point, the rational response for schools under pressure was to focus instructional time on the specific content and format of the tests rather than on broader learning. Test scores in some districts improved. Evidence of deeper understanding and transferable skills did not always follow. The metric had been optimized. The thing the metric was supposed to measure had not.
Healthcare Metrics Gone Wrong
Healthcare provides equally instructive examples. Hospital wait time targets, implemented in various national health systems to improve patient experience, produced a range of creative responses. Some hospitals addressed the underlying problem by improving throughput and staffing. Others addressed the metric by registering patients as seen before they had been properly assessed, keeping them in waiting rooms but technically on a different part of the clock. The target was met. The patient experience it was designed to measure did not always improve correspondingly.
Why Smart Organizations Keep Making This Mistake
Goodhart’s Law is not a failure of intelligence. The organizations and individuals who fall into its trap are frequently acting rationally given the incentive structures they face. The problem is a structural one: measurement is necessary, and imperfect proxies are often the best available tools. The challenge is not to stop measuring but to hold the relationship between the measure and the underlying goal with sufficient care that the decoupling does not go unnoticed.
Part of what makes this hard is that the decoupling tends to be gradual and initially invisible. A metric starts as a reasonable proxy, targets are introduced, behavior begins to optimize toward the metric, the gap between metric performance and actual performance slowly widens, and by the time the gap is obvious something has usually already gone wrong. The early signal is subtle: the metric is improving faster than the underlying reality seems to justify, or the metric is improving while other indicators that should correlate with it are not moving in the same direction.
Practical Responses to Goodhart’s Law
Acknowledging the law does not mean abandoning measurement. It means being more thoughtful about how metrics are designed, used, and periodically reviewed. A few principles follow directly from the framework.
Using multiple metrics that are difficult to simultaneously game is one of the more robust responses. If a single metric becomes a target, it can be optimized in isolation. If several metrics that represent different aspects of the underlying goal are all tracked, optimizing for one at the expense of the others becomes visible. The goal should be a portfolio of measures that are collectively harder to decouple from the reality they represent than any single measure would be.
Rotating or updating metrics periodically is another approach. A measure that people have not yet had time to optimize for retains more of its original signal. This is not always practical, but in domains where measurement is ongoing and the underlying goal is stable, refreshing the specific metrics used to track it can help maintain their validity.
Perhaps most importantly, treating metrics as evidence rather than as the thing itself is a mindset shift that the best managers, policymakers, and analysts tend to share. A metric tells you something. It does not tell you everything. Regularly asking whether improving performance on a measure is actually producing the outcomes you care about, and being genuinely open to the answer being no, is the most honest form of quality control available.
The nail factory, real or not, is a useful thing to keep somewhere accessible in the mind. Every time a new metric becomes a target, it is worth asking: what kind of nails are we going to end up with?