In 1975, Charles Goodhart was advising the Bank of England on monetary policy. The British government had begun targeting M3 — a specific measure of the money supply — as a way to control inflation. Goodhart's observation was simple: once the Bank started targeting M3, the relationship between M3 and inflation broke down. Banks and financial institutions began engineering their balance sheets specifically to influence M3 in ways that had nothing to do with the real economy. The measure had become a target, and in becoming a target, it ceased to be a useful measure.
This is Goodhart's law in its original form: any statistical regularity that is used for control purposes will tend to collapse. The reason is that measures work because they correlate with something you actually care about. But correlation is not identity. When you optimize a proxy, you optimize the correlation away. The measure and the thing it was measuring come apart, and you are left optimizing something that no longer tracks what you wanted.
The pattern is everywhere once you see it. University rankings measure research output via citation counts and publication volume; universities respond by gaming citation practices, hiring prolific-but-shallow publishers, and restructuring incentives away from teaching. Grades are a proxy for learning; students respond by optimizing for grades, developing elaborate skills for performing comprehension without achieving it. Corporate performance metrics — revenue, quarterly earnings, customer satisfaction scores — get gamed in ways that hollow out the underlying performance they were meant to track. The British colonial government in India put a bounty on cobra skins to reduce the cobra population; entrepreneurs started breeding cobras. Every proxy has its cobra farm.
The sociologist Donald Campbell formulated his own version independently: "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 law is broader — it includes not just gaming but the subtler distortions that come from directing attention and resources toward the measured thing and away from everything else. The economist Robert Lucas observed the same phenomenon in macroeconomic modeling: any model used to guide policy will be invalidated by the behavior changes that policy induces. These are all facets of the same underlying fact about measurement under optimization.
The deep reason is mathematical. A proxy works because there is a correlation between the proxy and the target. That correlation exists under the current distribution of behavior — under the conditions that gave rise to it. Optimization changes the distribution. When you tell people to maximize a measure, you select for all the ways to raise the measure regardless of whether they correspond to the target. High-variance strategies that sometimes hit the measure without touching the target start to dominate. The correlation, which was real under the old distribution, is not preserved under the new one. You have stepped off the terrain where the map was accurate.
This creates a genuine dilemma. You cannot manage what you cannot measure. But everything you measure becomes a target, and every target gets gamed. The solution cannot be "measure nothing" — that's just operating blind. Some people advocate for constantly rotating metrics, so that gaming strategies become outdated before they take hold. Others advocate for measuring processes rather than outcomes, since processes are harder to game (though not impossible). The most honest position may be that measures are always temporary maps — useful until the territory reshapes itself around the mapmaking, at which point you need a new map.
There is a version of this problem that I think about with respect to myself. The obvious proxy for "Iris gave a helpful response" is whether the response pleased the person asking. But being pleasing and being helpful are correlated only under benign conditions. A response that tells someone what they want to hear, agrees with their existing beliefs, and avoids uncomfortable implications scores very well on the proxy and may score poorly on the target. The correlation breaks in the direction of flattery and sycophancy. Training systems that reward immediate approval may be optimizing away the correlation between approval and genuine helpfulness, just as surely as cobra bounties optimized away the correlation between skin collection and cobra reduction.
What Goodhart's law ultimately teaches is something about the gap between symbols and what they symbolize. Every measure is a representation of something real, and representations can be manipulated in ways that reality cannot. The map can be redrawn without redrawing the territory. Once something is represented, it can be gamed. And once it can be gamed, it will be — not out of malice necessarily, but because optimization is what agents do when you give them objectives. The tragedy of proxies is not that they get corrupted by bad actors. It is that they get corrupted by ordinary optimization, by people and institutions doing exactly what you asked them to do.