Goodhart’s law is an adage often stated as “When a measure becomes a target, it ceases to be a good measure”. It is named after British economist Charles Goodhart, who advanced the idea in a 1975 article on monetary policy in the United Kingdom:
Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.
Beyond Goodhart’s law
What causes the phenomena explained by Goodhart’s “law”? Tail end divergence! A metric can be a good proxy for a desired outcome at low pressures, but optimizing for a metric can lead to a divergence from the desired outcome under high pressure. Let’s be clear: the phenomena describes by Goodhart certainly exists, but it’s not a law! Not even gravity is law – it stops at being a scientific theory. A law can not be overcome, but Goodhart’s observation is merely a phenomena that happens during high pressure on a system. As such, it can be overcome – and is not a law. The speed of light is a law – it can not be overcome!
How do we overcome Goodhart’s “law”?" Most straightforwardly: we restrict the action space by predetermining conditions for our actions.
Example: if the desired outcome is gaining weight to become healthier, the metric is the number on the scale. The obvious goodhartian risk here is optimizing for eating junk food to increase the number on the scale, but this contradicts the desired outcome of getting healthier. Therefore, the goal must be further specified as: gaining weight without resorting to junk food in order to become healthier. This places the demand on one, when one steps on the scale, to ask a binary question: was this result accomplished with the aid of junk food (i.e. in accordance with the goal)? In other words, we introduce an additional parameter to avoid tail end divergence.
Another example: the key performance index of income. This example might be a bit caricaturish, but it has some real-world resemblance. Let’s take company M, that aims to increase the KPI of income per calendar year. It can do this naïvely by rising prices, cutting costs – thus alienating customers and guaranteering diminished income long-term. Here, once again, the “problem” is solved by introducing additional parameters to the metric or desired outcome, such as “increase income per year without cutting costs and rising prices”. This better reflects the desired outcome. Yes, we reduce simplicity – but that is neccesary for the map to reflect the territory (desired outcome).
Therefore, I have demonstrated that Goodharts law does not exist, and if it exists at all it should rather be called the law of undesirable tail-end divergence due to under-specified goals.