Goodhart's Law in the Real World

Charles Goodhart was writing about monetary policy when he observed that any statistical regularity collapses once pressure is placed on it for control purposes. The line survived economics because everyone who manages anything eventually watches it happen to them.

The mechanism is simple. A metric is chosen because it correlates with something you care about and is easier to observe. The correlation held because nobody was trying to move the metric on purpose. The moment you attach a reward, you have created a population of people whose job is to break that correlation as efficiently as possible.

A proxy is honest right up until it becomes profitable.

Case one: the support queue that got faster and worse

A support organization tied bonuses to average time-to-close. Closure time fell by a third in two quarters, which read as an operational triumph. Reopen rate, which nobody was reporting to the executive team, rose by more than half. Tickets were being closed, not resolved, and the same customer contact was now generating two or three tickets instead of one.

Total handling cost per resolved issue went up while every dashboard showed improvement.

Case two: the hiring pipeline that filled and emptied

A talent team was measured on time-to-fill. The metric improved sharply. Ninety-day attrition among new hires roughly doubled, because the fastest way to fill a role is to lower the bar on the last interview, and the cost of that decision lands on a different team a quarter later.

This is a temporal mismatch as much as a measurement problem: the reward arrived in the same month as the decision, the cost arrived three months later on someone else's ledger.

Case three: the safety record that improved by silence

An operations group tied site bonuses to reported incident counts. Reported incidents fell. Severity of the incidents that did get reported climbed, which is the signature of a reporting problem rather than a safety improvement. The metric had turned a data-collection system into a liability for the people generating the data.

The fix: pair the metric and separate the definition

Two structural edits neutralize most Goodhart failures, and neither requires additional oversight.

First, pair the target with a counter-metric that only degrades when the shortcut is taken — closure time with reopen rate, time-to-fill with ninety-day survival, incident count with severity distribution and near-miss reports. A single number can be gamed cheaply; the pair usually cannot, because the same shortcut moves them in opposite directions.

Second, give definition control to someone downstream of the reward. When the team being measured also decides what counts, definition drift does the gaming quietly and no rule catches it.

• —Name the outcome the metric proxies, explicitly.
• —Attach one counter-metric that the cheapest shortcut would damage.
• —Move definition authority downstream of the incentive.
• —Review the pair together or the pairing is decorative.