Position · last reconsidered September 17, 2026

A measure can break in four ways

Garrabrant argues that a number stops tracking what it was meant to track in at least four different ways once people chase it, and each needs its own fix.

In our words

Scott Garrabrant takes Goodhart's lawGoodhart's lawWhen a measure becomes a target, it stops being a good measure: people learn to improve the number rather than what the number stood for., the observation that a measure stops working once people aim at it, and splits it into at least four failures. Picking the top scorers also picks the lucky ones. People raise the number by means that do nothing for the real goal. Push a number far enough and it leaves the range where it once matched the goal. Or someone games it on purpose, like a salesperson booking fake orders to hit a target. Each failure needs a different fix, so lumping them together hides what to do.

Scott GarrabrantRationalist community (LessWrong)2017

In their words

However, this is not a single phenomenon. I propose that there are (at least) four different mechanisms through which proxy measures break when you optimize for them.
Scott Garrabrant, Goodhart Taxonomy, opening paragraphRead the passage

Connected ideas

derived from

extends

related to

responds to