Goodhart's Law

In 2001, the British government introduced hospital waiting-time targets for the National Health Service, mandating that no patient should wait longer than four hours in an emergency department. Within months, hospitals began gaming the system. Some reclassified patients to stop the clock. Others moved patients into "clinical decision units" that technically counted as admissions, even though no treatment had begun. Ambulances were held in parking lots so patients would not enter the emergency department until the four-hour window was safely achievable. A 2010 Royal College of Emergency Medicine investigation found that the target had distorted clinical priorities: patients were treated by arrival time, not severity. The metric that was supposed to improve care had become the obstacle to providing it.

What Goodhart's Law Says

Goodhart's Law states that when a measure becomes a target, it ceases to be a good measure. The principle was first articulated by British economist Charles Goodhart in a 1975 paper on monetary policy, where he observed that any statistical regularity relied upon for policy purposes would tend to collapse once pressure was placed upon it. The idea was later generalized by anthropologist Marilyn Strathern into the pithier formulation most people encounter today: "When a measure becomes a target, it ceases to be a good measure."

This is NOT the same as saying that measurement is bad or that metrics are useless. Goodhart's Law is specifically about what happens when a proxy indicator -- chosen because it correlates with something you value -- is converted into a target that people are incentivized to hit. The correlation between the metric and the underlying goal exists precisely because no one was trying to manipulate it. The moment the metric becomes a target, people optimize for the metric itself rather than the goal it was designed to represent, and the correlation breaks down. The metric becomes noise, or worse, an active distortion.

Why the Correlation Breaks

The mechanism is rooted in what economists call the Lucas critique, named after Robert Lucas, who argued in 1976 that any observed statistical relationship will shift once agents change their behavior in response to policy based on that relationship. Goodhart was making essentially the same point in the monetary policy domain. The deeper psychological mechanism is straightforward: when you attach consequences -- bonuses, promotions, funding, survival -- to a number, rational agents will find the most efficient path to moving that number. Whether that path aligns with the original intent is a separate question entirely, and one that the incentive structure does not address.

This is not about dishonesty. The NHS administrators who gamed waiting-time targets were not villains. They were responding rationally to a system that rewarded one observable output while the actual goal -- quality of patient care -- remained largely unobservable. Economist Keith Hoskin and sociologist Richard Macve have argued that this dynamic is intrinsic to all measurement-based governance: the act of measuring changes the behavior of the measured. The measure and the thing it was meant to capture decouple precisely because optimization pressure is applied to the measure.

Real-World Manifestations

The Cobra Effect: Systemic gaming at national scale

The most vivid historical example is the cobra effect, which occurred during British colonial rule in Delhi. The colonial government, alarmed by venomous cobras, offered a bounty for every dead cobra brought to authorities. Initially, the program appeared to work. Then enterprising citizens began breeding cobras to collect the bounties. When the government discovered the perverse outcome and canceled the program, breeders released their now-worthless cobras into the streets, and the snake population ended up larger than before the intervention. The same pattern repeated when the French colonial government in Hanoi offered a bounty for rat tails. Rats with severed tails were soon spotted running through the streets -- entrepreneurs had been collecting tails and releasing the rats to breed more bounty-worthy offspring. In both cases, the metric (dead cobras, rat tails) was perfectly correlated with the goal (fewer dangerous animals) only until the metric was converted into a target.

Academic publishing: Personal-scale distortion

At the individual level, the "publish or perish" culture in academic research demonstrates Goodhart's Law with painful clarity. Researchers evaluated on publication count engage in "salami slicing" -- splitting one substantive study into three or four minimal publishable units. Citation counts, once a rough proxy for intellectual impact, are inflated through citation rings, strategic self-citation, and the production of review papers designed to accumulate references rather than advance knowledge. A 2018 study by John Ioannidis at Stanford found that a small number of researchers had accumulated extraordinary self-citation rates that dramatically inflated their apparent impact. The metric goes up; the quality of science does not. The researchers are not acting irrationally. They are responding to a system that has converted a proxy into a target.

The Multi-Metric Trap

A common organizational response to metric gaming is to add more metrics. If one KPI can be gamed, surely a dashboard of twenty will be harder to manipulate. But this creates its own pathology. People either optimize for whichever metric carries the highest personal payoff and ignore the rest, or they spread effort so thin across all metrics that nothing receives meaningful attention. Worse, multiple metrics frequently conflict with each other -- optimizing customer satisfaction scores may require longer call times, which violates the efficiency metric, which undermines the cost-per-interaction target. The result is not more accurate measurement but more elaborate gaming strategies. Complexity does not solve the fundamental problem; it makes the distortion harder to diagnose.

This is closely related to information asymmetry: the people being measured typically understand the system's loopholes and pressure points far better than the people designing the metrics. The measured have granular, real-time knowledge of how the system actually works; the measurers have a simplified model represented by their dashboard. This asymmetry ensures that gaming strategies will always outpace monitoring efforts in the long run.

Limitations and Failure Modes

Goodhart's Law is a powerful diagnostic tool, but it has its own boundaries and can be misapplied.

First, the principle can become an excuse for not measuring anything at all. Some organizational cultures, upon learning about Goodhart's Law, swing to the opposite extreme and abandon quantitative evaluation entirely in favor of vague qualitative assessments. This is a misreading. The law does not say measurement is useless -- it says that converting measures into targets introduces predictable distortions. The distinction matters. A thermometer is useful for understanding your health; optimizing your body temperature directly does not make you healthier. The problem is not the thermometer but the optimization.

Second, not all metric distortion is equally harmful. Some gaming is benign or even useful. Students "teaching to the test" in a well-designed examination are actually learning the material. Sales teams optimizing for a revenue target may be doing exactly what the business needs. Goodhart's Law applies most dangerously when the metric is a poor proxy for the actual goal -- when the gap between what you measure and what you value is wide.

Third, the law can be weaponized as a rhetorical device to dismiss any accountability system. "That metric will just be gamed" is technically true of every metric, which makes it a convenient argument for anyone who wants to avoid being measured. The productive response is not to abandon measurement but to design metrics that are harder to game, to rotate them regularly, and to supplement them with qualitative judgment.

Fourth, Goodhart's Law is sometimes confused with mere incompetence in metric design. Choosing a terrible metric and then watching it fail is not Goodhart's Law -- it is just bad measurement. The law specifically describes the phenomenon where a previously valid statistical relationship degrades because it was adopted as a target. The distinction matters for diagnosis: if a metric was never correlated with the goal, the problem is selection, not Goodhart.

Fifth, temporal dynamics matter. Metrics tend to be most useful early in their life cycle, before the agents being measured have figured out how to game them. The longer a metric persists unchanged, the more thoroughly it will be gamed. This means that even well-chosen metrics degrade over time, and any measurement system that does not regularly retire and replace its indicators will eventually be captured by the optimization it was designed to track.

Goodhart's Law connects to several other frameworks in ways that sharpen the diagnosis. Incentive structures are the transmission mechanism through which Goodhart's Law operates -- every time an organization attaches a reward or punishment to a metric, it creates the conditions for the measure-target decoupling to occur. Information asymmetry explains why gaming is so difficult to prevent: those being measured always know more about the system's exploitable features than those doing the measuring. Nudge theory offers a partial remedy -- rather than setting hard targets, designing choice architectures that make desired behaviors the path of least resistance can achieve alignment without creating the optimization pressure that triggers Goodhart distortions. Systems thinking provides the wider lens: Goodhart's Law is a specific instance of the general principle that intervening in a complex system produces unintended consequences, because the system adapts to the intervention in ways the intervener did not anticipate.

Building the Habit: The Pre-Mortem Test

The behavioral practice for Goodhart's Law awareness is the Pre-Mortem Test. Before committing to any metric as a target -- in your team, your organization, or your personal goals -- pause and ask: If someone optimized ruthlessly for this number and nothing else, what is the worst outcome that could result?

The internal experience is a shift from satisfaction ("we have a clear target") to unease ("wait -- what would gaming this look like?"). That unease is the signal that you are thinking about the gap between the metric and the goal it represents. The trigger situation is any moment when a number is about to be tied to consequences: a quarterly target, a bonus structure, a performance review criterion, a personal productivity goal. In each case, the pre-mortem reveals whether the metric is robust enough to survive optimization pressure or whether it will decouple from the underlying objective the moment pressure is applied.

Describing your desired outcome in a sentence, then checking whether your metric actually captures that sentence, is the simplest form of the test. The NHS wanted "patients receive timely, appropriate emergency care." The metric captured "patients are formally seen within four hours." The gap between those two statements is where Goodhart's Law lives.

Back to the Emergency Department

The NHS eventually reformed its approach. In 2019, after years of documented gaming and clinical distortion, NHS England announced it would replace the single four-hour target with a set of clinical outcome measures -- time to initial assessment, time to treatment for critical conditions, and patient-reported experience. The shift was not from measurement to no measurement but from a single gameable target to a more nuanced set of indicators designed to track what actually mattered: whether patients were getting the right care at the right time. Whether this new system will resist Goodhart's Law indefinitely is doubtful -- all metrics degrade under sustained optimization pressure. But the recognition that the original target had ceased to be a good measure, precisely because it had become a target, is Goodhart's Law understood and applied. The question is never whether to measure. It is whether you are measuring what you think you are measuring, and whether that will still be true once people know the number matters.