Goodhart's Law

You cannot manage what you cannot measure, the saying goes — so we measure. We can’t see “learning”, so we count test scores. We can’t see “health”, so we clock A&E wait-times. We can’t see “productivity”, so we log lines of code, calls closed, papers published, arrests made. Each number is a stand-in — a proxy — for something real we care about but can’t watch directly. And out in the wild, before anyone’s job depends on it, the proxy and the real thing move together: better teaching does tend to lift test scores, healthier hospitals do tend to have shorter waits. That correlation is exactly why we trust the number in the first place.

Then we do the fateful thing. We attach a stake to the proxy — a bonus, a ranking, a funding cut, a promotion — and tell people to make the number go up. And the number goes up. But the thing we actually wanted often doesn’t come with it, and sometimes goes down. Teachers drill test technique instead of teaching; hospitals leave ambulances queuing outside so the four-hour clock doesn’t start; officers arrest easy targets to hit a quota. The measure climbs while the reality it stood for quietly rots. That is Goodhart’s Law: when a measure becomes a target, it ceases to be a good measure. Named for the economist Charles Goodhart, who spotted it in monetary policy in 1975, and sharpened by the anthropologist Marilyn Strathern into the crisp form most people quote today.

This is an expert course, and it sits at the meeting point of three disciplines you’ve already visited. From incentives it takes the iron law that people optimise what is actually rewarded, not what you piously hoped they’d pursue — the engine that makes anyone bother to game a number at all. From feedback loops it takes the machinery of a self-reinforcing spiral: the proxy feeds back into decisions, effort floods toward it, and that very flood corrupts the signal the proxy once carried. From the principal–agent problem it takes the wall of hidden knowledge — the measured party always knows more about how the number was made than the measurer does, and profits from the gap. Goodhart’s Law is what you get when those three models collide: a special, brutal case where the rewarded proxy and the real goal, correlated in the wild, come apart the instant you push on them.

Across the course we build the model floor to ceiling. We start with proxy versus goal — why we ever settle for a stand-in, and what makes a proxy trustworthy until it isn’t. Then the mechanism of decoupling under pressure: Campbell’s Law, and the four flavours of Goodhart (regressional, extremal, causal, adversarial) kept intuitive, so you can see how the seams split. We tour the flagship disasters — the Soviet nail factory that made one seventy-tonne nail to hit its weight quota, the colonial cobra bounty that bred cobras, teaching to the test, sales sandbagging, clickbait, A&E trolley-waits, arrest quotas, salami-sliced science, the risk metrics gamed to nothing before 2008 — because the same skeleton wears a hundred costumes. We show precisely why it is a special case of incentives plus principal–agent, wired into a feedback loop that eats the metric’s own validity. And we stay ruthlessly honest about where the model lies to you: it is not an argument against measuring (flying blind is worse), not every measure collapses, and quoting “Goodhart!” to dodge all accountability is its own lazy failure. We finish with the fixes that actually work — baskets of hard-to-game metrics, targets held loosely, humans kept in the loop, rewarding outcomes over proxies, and expecting to rotate a metric as it decays.

Throughout, you’ll have a Goodhart pressure dial to play with — a crowd of people whose real quality and measured proxy start out marching in lockstep, until you turn up the stakes and watch the proxy rocket, the real value sag, and the correlation between them crumble to nothing before your eyes. By the end you’ll never look at a KPI, a league table, or a “we hit our numbers” the same way again. You’ll ask the sharp question instead: this number is going up — but is the thing it was supposed to stand for going up with it, or is somebody just very good at making the number go up?