Goodhart's Law

What Goodhart’s Law means

Goodhart’s Law states that when a measure becomes a target, it ceases to be a good measure. The moment you attach incentives or consequences to a metric, people begin optimising for the metric itself rather than for the underlying reality it was designed to capture. The metric still moves - often in the desired direction - but the thing it was supposed to represent stops improving, or actively deteriorates.

The principle is named after British economist Charles Goodhart, who originally formulated it in 1975 in the context of monetary policy. His observation was that statistical relationships used to guide policy broke down the moment they were used for that purpose. The act of targeting a measure changed the behaviour of the system it was measuring, invalidating the relationship that made the measure useful in the first place.

The anthropologist Marilyn Strathern later generalised the idea into its most quoted form: “When a measure becomes a target, it ceases to be a good measure.” This version captures the universal principle: metrics are useful as observations of a system, but the moment they become objectives within the system, they distort the very behaviour they were tracking.

How Goodhart’s Law works

The corruption of indicators

Every metric is an imperfect proxy for the thing you care about. Hospital waiting times are a proxy for quality of care. Exam results are a proxy for learning. Website traffic is a proxy for audience engagement. Revenue is a proxy for business value. These proxies are useful precisely because they correlate with the underlying reality - at least, they do until you start targeting them.

Once a metric becomes a target, people find ways to improve the metric that don’t improve the underlying reality. Hospitals reschedule patients to avoid breaching waiting time thresholds. Schools focus teaching on tested material at the expense of broader education. Websites publish clickbait that inflates traffic but degrades the reader’s experience. The metric improves. The reality doesn’t. Sometimes the reality gets worse, because the energy spent gaming the metric is energy diverted from the actual work.

This is map is not the territory applied to organisational life. The metric is a map of reality. It’s useful for navigation. But optimising the map rather than the territory it represents is a category error - and it’s one that organisations commit systematically.

The incentive trap

Goodhart’s Law isn’t caused by dishonesty. It’s caused by incentives. When people are rewarded for hitting a number, they will - rationally and predictably - find the most efficient way to hit that number. If the most efficient route to the number happens to align with the intended outcome, the system works. If it doesn’t - and it increasingly doesn’t as the metric becomes the focus - the system breaks.

The trap is that the breakdown is often invisible to anyone looking only at the numbers. The dashboard looks healthy. The targets are being met. But the lived reality - the quality of care, the depth of learning, the value of the product - has quietly decoupled from the metrics that are supposed to represent it. By the time the gap becomes visible, the damage is often substantial.

Compounding distortion

Goodhart’s Law tends to get worse over time through feedback loops. Once people start optimising for a metric, the metric becomes less reliable. Because it’s less reliable, decision-makers add more metrics, or tighten the targets, or introduce new incentives. These additional controls create additional opportunities for gaming. The system becomes increasingly complex, increasingly gamed, and increasingly disconnected from its original purpose.

This is a specific instance of a broader systems principle: attempts to control complex systems through simple measures tend to generate unintended consequences that require further interventions, which generate further unintended consequences, in an escalating cycle. The solution to a metric problem is rarely another metric.

Goodhart’s Law in everyday life

Education

Education is one of the most visible arenas for Goodhart’s Law. When exam results become the primary measure of school quality, schools predictably redirect resources toward exam preparation. Teaching to the test improves test scores while narrowing the curriculum and reducing the depth of genuine learning. Students learn to pass exams rather than to understand subjects. The metric rises. The education arguably declines.

League tables amplify the effect. Schools that serve disadvantaged communities face pressure to match the outcomes of schools serving affluent ones, leading to strategies that optimise for measurable outputs - exam passes, attendance figures, exclusion rates - rather than for the messier, harder-to-measure reality of educational quality. In extreme cases, schools have been caught manipulating results directly, but the more insidious damage happens through the perfectly legal reallocation of time and resources toward whatever is being measured.

Healthcare

Healthcare systems worldwide illustrate Goodhart’s Law with depressing regularity. When hospitals are measured on waiting times, they find ways to reduce waiting times that don’t improve care - reclassifying patients, gaming referral pathways, starting the clock at different points. When surgeons are measured on mortality rates, some avoid operating on high-risk patients who might benefit most from surgery. The metric that was supposed to improve patient outcomes instead creates incentives to avoid the patients with the worst outcomes.

The NHS in the UK has experienced multiple cycles of this pattern: targets are introduced to address genuine problems, the targets are met through gaming, the gaming is discovered, new targets are introduced, and the cycle repeats. The underlying challenge - that healthcare quality is multidimensional and resistant to simple measurement - remains unresolved.

Technology and social media

Goodhart’s Law runs rampant in the technology industry. Social media platforms optimise for engagement metrics - time spent, clicks, shares, comments. Engagement is a proxy for providing value to users. But once engagement becomes the target, algorithms learn that outrage, conflict, and misinformation generate more engagement than nuance, accuracy, or calm. The metric soars. The user experience deteriorates. The attention economy is, in many ways, a system-wide expression of Goodhart’s Law: the metrics that platforms optimise for have become decoupled from the outcomes that would benefit users.

In software development, similar patterns emerge. When developers are measured on lines of code, they write verbose code. When they’re measured on number of features shipped, they ship features regardless of quality. When they’re measured on bug counts, they reclassify bugs as features. The measure becomes the target, and the target becomes the game.

Business and management

Corporate environments are fertile ground for Goodhart’s Law. Sales teams measured purely on revenue may close deals that damage long-term relationships. Customer service teams measured on call resolution time may rush customers off the phone without solving their problems. Manufacturing teams measured on output may sacrifice quality for volume. In each case, the metric faithfully records improvement while the reality it was supposed to capture quietly degrades.

The quarterly earnings cycle is arguably the largest-scale example of Goodhart’s Law in modern capitalism. When a company’s value is judged primarily by short-term financial metrics, management optimises for those metrics - cutting investment, deferring maintenance, laying off staff, buying back shares - at the expense of long-term health. The metric that was supposed to represent corporate performance instead drives behaviour that undermines it.

Policing and criminal justice

Crime statistics are notoriously vulnerable to Goodhart’s Law. When police forces are measured on crime rates, they have incentives to reclassify crimes, discourage reporting, or focus resources on easy-to-solve cases rather than serious but complex ones. When they’re measured on arrest numbers, they have incentives to make arrests regardless of whether they lead to convictions. When they’re measured on stop-and-search figures, they have incentives to conduct searches regardless of whether they’re justified.

The metric tells a story. The story may have very little to do with whether communities are safer or whether justice is being served.

How to work with Goodhart’s Law

Measure many things

The more dimensions you measure, the harder it is to game any single one. A school measured on exam results, student wellbeing, breadth of curriculum, and long-term outcomes is harder to game than one measured on exam results alone. Multiple metrics don’t eliminate Goodhart’s Law, but they raise the cost of gaming to the point where actually doing the work becomes the easier option.

Separate measurement from incentives

Metrics are most reliable when they’re observed rather than targeted. Using data to understand a system is different from using data to reward or punish people within that system. Where possible, maintain a firewall between the metrics you use for insight and the metrics you attach consequences to. The moment a metric carries a reward, assume it will be gamed.

Use second-order thinking

Before introducing a target, ask: if people optimise for this number, what will they do? Not what would I like them to do, but what will they actually do? What’s the easiest way to hit this target without doing the intended work? If you can see a gaming strategy, so can everyone else. Design the metric to make the gaming strategy harder than the genuine strategy.

Accept irreducible complexity

Some things that matter most - learning, health, justice, trust, quality of life - are inherently difficult to measure. Goodhart’s Law is most damaging when it’s applied to domains that resist simple quantification. Accepting that some important things can’t be fully captured by metrics isn’t a failure of measurement. It’s an honest acknowledgement of complexity - and it’s the starting point for managing those domains wisely rather than just counting them badly.

Goodhart’s Law doesn’t mean measurement is useless. It means measurement is powerful - powerful enough to reshape the system it’s applied to. Understanding that power, and designing measurement systems that account for it, is one of the most important challenges in any organisation, institution, or society that relies on data to make decisions. Which, in the modern world, is all of them.