There’s an old story — probably apocryphal, but too good not to teach with. A Soviet nail factory is set a production target measured in tonnes. So it makes one gigantic, useless nail weighing seventy tonnes and hits the quota exactly. Horrified, the planners switch the target to the number of nails. So the factory stamps out millions of tiny, useless tacks. Both times the factory hit its number perfectly. Both times it produced nothing anyone could build with.
That is Goodhart’s Law in one grim joke: when a measure becomes a target, it ceases to be a good measure. The factory was never trying to make nails — it was trying to make the number. And once making the number is the whole game, the number and the thing it was supposed to stand for peel apart.
What this course gives you
Goodhart’s Law is one of those models that, once you see it, you cannot unsee. It’s hiding inside every KPI, every league table, every “we hit our targets”, every algorithm optimising for engagement. By the end you’ll be able to name the mechanism precisely, recognise it in a dozen disguises, and — crucially — know what to actually do instead of throwing your hands up. Here’s the map:
- Proxy vs goal — why we measure stand-ins at all, and what makes a proxy trustworthy right up until it isn’t.
- How the seams split — the mechanism of decoupling under pressure: Campbell’s Law and the four flavours of Goodhart.
- A hundred costumes — the flagship disasters, from cobra bounties to clickbait to the 2008 risk models, all wearing the same skeleton.
- Why it’s a special case — how Goodhart falls straight out of incentives + feedback loops + the principal–agent problem.
- What to do about it — the fixes that work, and the ways quoting “Goodhart!” becomes its own lazy excuse.
The one-sentence version
A metric is only ever a proxy for what you truly care about; the moment you apply hard optimisation pressure to the proxy, the correlation you were relying on breaks — and you get the number without the substance.
Before you read — take a guess
Before we start — take a guess. What does Goodhart's Law claim?
A quick taste: the honest number that goes bad
Picture a school where, quietly, good teaching and good test scores rise together. Nobody’s gaming anything, so the scores are an honest read on learning — a teacher with better scores really does teach better. The number is trustworthy.
Now make those scores decide teacher pay, school funding, and headlines. Suddenly the smart move isn’t to teach better — it’s to raise the score: drill last year’s paper, coach exam technique, quietly counsel weak pupils to stay home on test day. Scores climb. Learning doesn’t. The very act of rewarding the number destroyed its ability to tell you anything about the thing you cared about.
Where we're headed
That school is the whole law in miniature. “Scores track learning when nobody’s watching” is why we trust a proxy. “Reward the scores and they detach from learning” is Goodhart’s Law biting. The gap between the two — the number soaring while the substance sags — is exactly what we’ll learn to see, explain, and defend against.
In the school example, why did test scores stop reflecting learning once they decided pay and funding?
How to use this course
Every lesson opens with a quick guess, teaches the idea through a concrete case, and checks that it stuck. Don’t skip the guesses — trying to answer before you know is one of the most reliable ways to remember. There’s a Goodhart pressure dial you’ll meet in the next lesson that lets you watch a proxy detach from reality as you crank the stakes, and light exercises throughout to keep the ideas honest.
When you’ve finished the five teaching lessons, a graded final exam pulls it all together. It’s one-way — once you submit an answer it’s locked — so treat the practice questions along the way as exactly that: practice.
One thing to unlearn as you go
“We hit our numbers” feels like success and often is. But this course will leave you unable to hear that phrase without a second question forming automatically: did the thing the number stood for improve too — or did someone just get very good at moving the number? That reflex is the whole skill.
Ready? The next lesson starts where every case of Goodhart’s Law starts: with the innocent, necessary decision to measure a proxy for something we can’t see directly.