Goodhart's Law and the Death of Common Sense

In the Soviet Union, central planners set production targets for nail factories. The target was measured in weight — tonnes of nails produced per quarter. The factories responded rationally. They produced enormous nails. Railroad spikes the size of your forearm. The targets were met. The country had no usable nails.

The planners, recognising the problem, switched the metric. The new target was measured in quantity — number of nails produced per quarter. The factories responded rationally. They produced millions of tiny, useless nails — essentially metal splinters. The targets were met. The country still had no usable nails.

This story — possibly apocryphal but pedagogically perfect — is the purest illustration of a principle that governs far more of modern life than anyone is comfortable admitting. In 1975, British economist Charles Goodhart stated it formally:

"When a measure becomes a target, it ceases to be a good measure."

The moment you tell people they'll be judged by a number, they optimise for the number — not for the thing the number was supposed to represent. And the harder you push the number, the wider the gap between the metric and the reality it was designed to capture.

The Metric That Eats Itself

Goodhart's Law isn't about dishonesty. The Soviet factories weren't trying to sabotage production. They were responding logically to the incentive structure. The problem wasn't in the response — it was in the target.

Every metric is a simplification. A number that tries to capture a complex, multi-dimensional reality. Revenue is a simplification of business health. Test scores are a simplification of student learning. Body weight is a simplification of physical fitness. Crime rates are a simplification of public safety. Each metric captures one dimension of a thing that has dozens of dimensions.

When you measure without targeting, the metric gives you useful information. Revenue tells you something true about the business. Test scores tell you something true about the student. The metric is a window into reality.

When you turn that metric into a target — when you say "revenue must be $X" or "test scores must reach Y" — the metric stops being a window and becomes a mirror. People start managing the metric instead of managing the reality. The number improves. The underlying thing it was measuring may not.

Campbell's Law

Sociologist Donald Campbell stated a stronger version in 1976: "The more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor."

Campbell's version is darker. It says the corruption isn't just incidental — it's proportional to the stakes. The more important the metric, the more it gets gamed. The more it gets gamed, the less useful it becomes. The less useful it becomes, the more the system relies on it — because by now, the metric has replaced the judgement it was supposed to inform.

The Modern Nail Factory

You don't need to visit the Soviet Union to see Goodhart's Law in action. You just need to look at your own workplace.

The Call Centre

Customer service is measured by "average handle time" — how quickly agents resolve calls. The target: keep handle time under four minutes. The rational response: agents rush through calls, transfer problems to other departments, and discourage customers from explaining complex issues. Handle time drops. Customer satisfaction craters. But the dashboard looks great.

The metric was designed to capture efficiency. When it became a target, it optimised for speed at the expense of the thing efficiency was supposed to serve: actually solving the customer's problem.

The Hospital

Hospital quality is measured by mortality rates — the percentage of patients who die. Logical enough. But when mortality rates became public metrics that affected hospital funding and rankings, something predictable happened: hospitals began avoiding high-risk patients. Why operate on someone with a 30% chance of dying when their death will worsen your metric? Better to transfer them to another hospital.

The metric was designed to measure quality of care. When it became a target, it incentivised avoiding the patients who needed the most care. The numbers improved. Patient outcomes — for the most vulnerable patients — got worse.

The School

Standardised testing was designed to measure student learning. When test scores became targets — tied to school funding, teacher evaluations, and administrative careers — the response was predictable. Schools began "teaching to the test." Subjects not on the test — art, music, physical education, critical thinking — were de-prioritised. Within tested subjects, instruction narrowed to the specific question formats and content areas that appeared on the exam.

Test scores rose. Actual learning — the breadth, depth, and applicability of knowledge — is harder to measure, but every teacher in the system will tell you it declined. The metric ate the thing it was supposed to measure.

Your Fitness Tracker

You set a goal: 10,000 steps per day. Within a week, you're pacing your living room at 11pm to hit the number. You're not healthier — you're gaming a metric. The step count was supposed to be a proxy for "move your body more." The moment it became a target, you started optimising for steps instead of health. You take the stairs not because it's good for you but because your wrist vibrates. You've Goodharted yourself.

Goodhart's Law is not about bad metrics. It's about what happens to good metrics when you attach consequences to them. The metric is useful precisely because it's a simplification. But useful simplifications become dangerous when people start optimising for the simplification instead of the reality.

The Cobra Problem (Again)

Goodhart's Law and the Cobra Effect are cousins. Both describe situations where a well-intentioned intervention — a bounty, a target, a metric — produces the opposite of its intended effect by changing the behaviour of the people inside the system.

The distinction is that the Cobra Effect is about interventions (doing something specific to solve a problem) while Goodhart's Law is about measurement (tracking a number to monitor progress). But the underlying mechanism is identical: when you change the incentive structure, you change the behaviour, and the behaviour changes in whatever direction the incentive points — which may not be the direction you wanted.

The Vietnam Body Count

During the Vietnam War, the U.S. military used enemy body counts as the primary metric of success. The logic seemed sound: if we're killing more of them than they're killing of us, we must be winning.

Officers whose promotions depended on body counts had every incentive to inflate them. Civilian casualties were counted as enemy combatants. The same body was sometimes counted multiple times. Units conducted operations designed to maximise body count rather than achieve strategic objectives.

The metric said America was winning. America was not winning. But the gap between the metric and reality was invisible to everyone who was looking at the metric — which was everyone, because the metric was the only thing being measured.

Designing Metrics That Survive Their Own Success

If every metric degrades when it becomes a target, does that mean measurement is futile? No. It means measurement requires design — specifically, design that anticipates Goodhart's Law and builds defences against it.

Use Multiple Metrics in Tension

A single metric is a single dimension. It will be gamed along that dimension. Multiple metrics that pull in different directions are harder to game because optimising one comes at the cost of another.

The call centre should measure handle time AND customer satisfaction AND first-call resolution rate. A short call that doesn't solve the problem scores well on one metric and poorly on the others. An agent gaming the system has to game all three simultaneously — which is much harder, and which starts to approximate actually doing the job well.

When your metrics are in tension with each other, gaming becomes equivalent to performance. The agent who satisfies all three metrics is, by definition, handling calls quickly, solving problems, and leaving customers happy. The metrics become self-correcting.

Measure Outcomes, Not Outputs

The Soviet nail factory measured outputs — tonnes produced, nails produced. What it should have measured was outcomes — buildings constructed, structures reinforced, customer orders fulfilled. Outputs are activities. Outcomes are results. The gap between them is where Goodhart lives.

A school should measure what students can do six months after the test, not what they can recall during it. A hospital should measure patient health outcomes over time, not mortality in the building. A company should measure customer lifetime value, not quarterly revenue. Each of these is harder to measure. That's exactly the point — the metrics most resistant to gaming are the ones closest to the real thing you care about.

Rotate and Retire

Metrics have a half-life. The longer a metric is in use as a target, the more the system learns to game it. The most Goodhart-resistant organisations rotate their metrics — changing what's measured before the gaming behaviour fully adapts.

Intel's Andy Grove was famous for using "paired metrics" that he rotated quarterly. One quarter, the focus was speed of delivery. Next quarter, quality of delivery. The system never had time to fully optimise for one dimension before the dimension changed.

Preserve Human Judgement

The deepest defence against Goodhart's Law is the simplest: don't let the metric replace the judgement. Use the metric as input to a human decision, not as the decision itself.

A teacher looking at test scores and thinking "this tells me something about where my students are" is using the metric well. A bureaucrat looking at test scores and thinking "this school must improve by 5% or lose funding" has let the metric replace the judgement. The first is measurement. The second is Goodhart's Law loading its weapon.

The Meta-Lesson

There's something philosophically unsettling about Goodhart's Law. It says that the act of measuring changes the thing being measured — not because the measurement is inaccurate, but because the measurement creates incentives, and incentives change behaviour. The observer affects the observed. The map reshapes the territory.

This is, in a sense, a social-scientific version of the Heisenberg Uncertainty Principle. In physics, you can't observe a particle without affecting it. In organisations, you can't measure behaviour without changing it. The measurement is never neutral. It's always an intervention.

The Soviet planners weren't wrong to want more nails. The call centre managers weren't wrong to want faster calls. The school administrators weren't wrong to want better learning. They were wrong about something more fundamental: they assumed they could point a metric at a complex system and the system would move in the direction of the metric.

Systems don't work that way. Systems respond to metrics the way water responds to a dam — by finding every other way around it.

Every improving dashboard hides the same question: "What is being optimised that isn't being measured?" Somewhere behind the metric that's climbing, there's a dimension that's degrading -- one that nobody is watching, because everyone is watching the number that's going up.

The nail factory hit its targets every quarter. The country ran out of nails.