Goodhart's Law: When the Measure Becomes the Target
Why any measure, once it becomes the goal people are judged by, stops accurately measuring the thing it was built to track.
A school is struggling, so the ministry sets a clear target: raise the pass rate. Within two years, the pass rate climbs sharply. Within three, everyone quietly knows why. Weaker students are discouraged from sitting the exam at all, teachers drill exam technique instead of the subject, and the number goes up while the actual learning, the thing the number was supposed to stand in for, barely moves.
This is Goodhart's Law, and once you have seen it once, you start seeing it everywhere.
Where It Comes From
The idea is named after Charles Goodhart, a British economist who made the observation in 1975 while studying monetary policy. His original point was narrow and technical: once a central bank starts targeting a particular financial indicator, the indicator stops behaving the way it used to, because people now have a reason to move it on purpose.
The anthropologist Marilyn Strathern later restated it in the plain sentence most people actually quote: when a measure becomes a target, it stops being a good measure. That restatement is why the idea escaped economics altogether and became one of the more useful mental models for understanding organizations, schools, and systems of almost any kind.
The Mechanism
Understand what is actually happening underneath this. A measure works well as long as it is a faithful proxy, a stand-in that tracks the real thing you care about without anyone paying direct attention to the proxy itself. The moment you turn that proxy into a target, something people are paid on, graded on, or praised for, you give them a reason to optimize the proxy directly. And a number can usually be moved more easily than the reality it was supposed to represent.
The gap between the measure and the real outcome was always there. Turning the measure into a target is what gives people a reason to walk straight into that gap and live there.
Where It Shows Up
Education. Pass rates, standardized test scores, and graduation rates all face this. Teach to the test, and the test score rises while actual understanding can stay flat or fall.
Business and sales. A target of units sold produces discounting and padded deals near quarter-end. A target of calls resolved per hour produces calls that get closed fast, not calls that get solved well.
Policing. Arrest quotas or ticket quotas produce more arrests and tickets. They do not reliably produce less crime, and they can produce the appearance of enforcement activity where none was actually needed.
Healthcare. A hospital measured purely on patient wait times can hit the target by reclassifying cases, adjusting how patients are logged, or rushing triage, none of which necessarily makes care faster or better.
Software and engineering. Lines of code written, tickets closed, or commits per week, each one is a weak proxy for actual engineering output, and each one can be inflated without producing more real value.
In every case, the pattern repeats: the number the organization chose to watch starts moving in the right direction, while the thing that number was supposed to represent quietly stops.
Why This Matters More Than It Seems
This is not simply a warning about dishonest people gaming a system. Most of the time, nobody is cheating. People are responding rationally to the incentive they were actually given, which is rarely the incentive the organization meant to give them. The failure sits in the design of the measure, not in the character of the people optimizing it.
This connects directly to first principles thinking. The real goal, whether that is "students understand the subject" or "customers get their problem solved," is the irreducible thing that actually matters. The measure, pass rate, average call time, is a convenient stand-in built on top of it, and convenient stand-ins are exactly the kind of inherited assumption first principles thinking asks you to examine rather than accept.
It also connects to falsifiability. A good measure should be one where you can ask, honestly, what would make this number rise while the real outcome falls. If you can picture that scenario clearly, and most of the time you can, you already know where the measure is going to be gamed before anyone has gamed it.
Guarding Against It
Watch for a gap opening between the proxy and the real outcome. If pass rates are rising but employers or downstream exams suggest graduates understand less, the proxy and the outcome have separated, and the proxy is the one lying.
Use more than one measure, chosen so they are hard to game together. A single number is always easiest to optimize in isolation. A call centre measured on both speed and a customer follow-up satisfaction score has a much harder time gaming both at once than gaming either alone.
Ask what behaviour the measure actually rewards, before adopting it. Before setting a target, picture the laziest, most literal way someone could hit that number without producing the outcome you want. If that path is easy to picture, the measure needs rethinking before it needs rolling out.
Revisit the measure periodically, on purpose. A measure that was faithful when it was introduced can stop being faithful once people have had a year to learn exactly how to move it without moving the reality underneath.
Questions to Ask
- What is the real outcome I actually care about, underneath this number?
- Could this number rise while that real outcome gets worse?
- What would the laziest way to hit this target look like, and am I comfortable with it?
The Actual Lesson
Goodhart's Law is not an argument against measuring anything. Organizations that refuse to measure drift blindly, which is its own failure. The lesson is narrower, and sharper: the moment a measure becomes a target that people are judged by, treat it with suspicion proportional to how much is riding on it, and keep checking whether it still points at the thing you actually meant to track.
The number was never the point. It was only ever supposed to be a window onto the point. Watch for the moment someone starts polishing the window instead of looking through it.