A measure that becomes a target ceases to be a good measure
Aliases: measure becomes a target · Goodhart · targetized metric
What it is
Goodhart's law says that when a measure becomes a target, it ceases to be a good measure. Session length once indicated involvement; once it is written as the success criterion, autoplay and infinite scroll lift it, and the old relationship between length and involvement breaks. The law is about a change in the status of a measure, not about a failure of character. The information a metric carries while it is not being optimized comes from a correlation with the goal that has not yet been put under pressure. After it is put under pressure, that correlation itself is what the optimization rewrites.
Why it happens
A metric can measure because, in the then-current distribution of behavior, it co-varies with the goal. Making it the target applies selection pressure to that co-variation: strategies that lift the metric without lifting the goal get paid; strategies that lift the goal without lifting the metric get dropped. The distribution therefore shifts from “natural co-variation” to “acting for the metric,” and the original correlation is hollowed out. Interfaces are especially good at offering shortcuts that never pass through the goal: enlarge a button to raise clicks, make confirmation skippable to cut time, pop the survey at a just-completed success to raise the score. Loss of meaning is asymptotic: early on the metric is still roughly usable; as shortcuts are found and copied, residual information falls. By the time the number looks best, it often represents the goal least.
Studying it
Compare the same metric’s correlation with an independent criterion in an “observe only” period and in a “written as a target” period; a drop is the law at work. The criterion should sit close to the goal and stay out of rewards—unpublished diary task success, coded spontaneous complaints in support. A lever audit also works: list interface strategies that appeared after the metric was pressured, and judge whether they bypass the goal. In the lab, an incentive manipulation can tell half the participants that a behavior will be scored, and watch whether the scored item decouples from actual task quality.
Where it stops holding
The law does not say that any targetization instantly ruins a metric. Hollow-out is slower when the window is short, the action space is small, or the metric nearly is the goal (a mandated completion rate). Treating the law as “so set no targets” leaves the organization without direction. What is required is to admit that targetization consumes measurement, not to imagine a number that remains valid forever while never being optimized. Continuing to announce success with an already-hollowed metric is a second misdirection.
Applying it
- Before a metric enters planning, ask which interface shortcuts could lift it without passing through the goal.
- Keep at least one criterion out of the reward, to watch whether the headline’s correlation with the goal is falling.
- When headline and criterion decouple, stop announcing success with the headline; close the shortcut or change the metric first.
- Do not dodge the law by folding weights into a composite: once the composite is the target, it hollows out the same way.