Goodhart's Law

A school wants students who can read. It starts paying, ranking, and punishing on a reading-score cutoff. Within a year the cutoff rises, and the students are not clearly better readers. Class time has moved to the items on the test. That is Goodhart's law in ordinary clothes. The number was a clue about the goal. The moment the number became the thing people were graded on, they had a reason to move the number without moving the goal.

Charles Goodhart noticed a version of this in monetary policy: a regularity you lean on for control tends to break once people adapt to the control. The line that stuck is plainer. When a measure becomes a target, it stops being a good measure. Teachers, nurses, and sales teams are often perfectly aware of the gap. They are answering the score you published. Teaching the items on the test, closing a ticket so you can reopen it, starting the wait-time clock after the patient sits down: that is how a decent count goes hollow.

A metric can still be worth keeping. Metric fixation, the neighbor on this site, is the earlier step: if you cannot count it, it barely exists. Goodhart's law is the step after that, when a count you chose on purpose becomes the target and then stops meaning the goal. The moment pay, rank, or fear attaches to the number, you need a second way to tell whether the goal is still in the room.

• "Wait times fell because we started the clock after the patient sat down."
• "Ticket volume is the KPI, so we close tickets and open new ones."
• "The engagement goal is the product now. The class can wait."
• "Crime is down on the dashboard we are scored on. Reports got harder to file."
• "We hired to the rubric. The rubric is what we know how to game."
• "Length of stay dropped. Readmissions do not count in my number."
• "If it is not in the quarterly metric, it is not my job this quarter."

The reading cutoff: the school is praised because the score jumped. Class time that used to be a book is now items that look like the test. A child who can ace the items still stalls on a paragraph that was not in the packet. The packet is what the score can see.

The queue: wait time hits target the month the clock starts after people sit down, not when they walk in. The dashboard is green. The bench by the door is the same bench.

The streak: you walk to the corner and back so the app will not break a chain. The hill you started the app in order to climb is still unclimbed. The chain feels like the habit.

Organizations get better at the scoreboard and worse at the reason the scoreboard was built. People doing the real work spend judgment on defending the number. Unmeasured goods, care, learning, safety margins, quiet maintenance, thin out because they do not pay. Trust erodes twice: outsiders notice the game, and insiders learn that naive honesty is how you miss target.

There is a personal version. A step count, a streak, a follower total, or a savings rate can start as a mirror and end as the life. You hit the number on a day you did not do the thing the number was a proxy for.

Numbers compress a messy goal into something a ministry, a manager, or an app can rank. Compression is useful until the compressed version is rewarded. People optimize what is rewarded. They have more information about loopholes than the designer of the metric does. A single indicator also hides tradeoffs: speed against accuracy, volume against severity, activity against outcome. Under pressure, the tradeoff is resolved in favor of whatever the formula sees.

Goodhart's 1975 observation came from monetary management, not from schools or software. Strathern (1997), writing about audit in British universities, stated the line that traveled: when a measure becomes a target, it ceases to be a good measure. She was describing assessment that creates the performance it then records. Campbell (1979) warned that quantitative social indicators used for decisions become open to corruption and distort the processes they monitor. He used the language of social programs, tests, and crime statistics.

None of these is a laboratory effect size. They are repeated observations about incentives. Use them as a design warning. A metric that still correlates with the goal after people have a reason to manipulate it is doing harder work than a metric that correlated before anyone was scored on it.

• You can raise the number without being able to explain how the goal improved.
• Colleagues joke about the loophole and still use it.
• Work that the formula ignores has become embarrassing to spend time on.
• A new metric appears every time the old one is "solved."
• You feel virtuous on a high-score day that you would not defend in plain words.

Pair the target with a check the target does not control: a sample of the actual work, a complaint that is not in the formula, a delayed outcome, a visit. Rotate or widen measures so a single loophole cannot be the whole job. Ask, before you publish a number as a goal, "If someone hits this and misses the point, how will we know?"

Personally, keep one description of the goal that is not the metric. "I want to be able to walk up the hill" can sit next to a step count. If the count is high and the hill is still impossible, the count has become the target.

Should we stop measuring?

No. Unmeasured work can hide failure too. The move is to stop pretending a rewarded measure is still an innocent instrument. Measure, and also audit the gap between the measure and the point.

Is gaming always bad faith?

Often it is local sense. If your shift, school, or bonus depends on a narrow number, protecting that number is what the system asked for. Blame the incentive before you blame the person who adapted to it. Then change the incentive.

When the number moves, ask what else moved. The goal is the thing you would still want if the dashboard went dark.

The cutoff

The queue

The streak

Practice this pattern in the Reframing App - capture the trigger, label it (like Goodhart's Law), check evidence, and write a more balanced thought.