Goodhart's Law: How to Break Every System

"When a measure becomes a target, it ceases to be a good measure." Charles Goodhart, 1975

Goodhart's Law is one of those ideas that feels obvious once you've seen it and then keeps surprising you for the rest of your life. It explains why so many well-meaning systems gradually warp into performance art, and why metrics are a great servant but a catastrophic master.

The original phrasing, from the British economist Charles Goodhart in 1975, is a little drier:

"Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes."

The simpler version: when a measure becomes a target, it stops being a good measure.

How it goes off the rails

You pick a metric because you want to know how something is going. Website engagement, test scores, bananas shipped per quarter. At first the metric is useful. It tells you, roughly, the thing you cared about. So you start optimizing for it directly. And then something shifts. The metric stops being a signal. It becomes the game.

And the game gets weird.

A school district decides to focus on test scores. Teachers teach to the test. Scores go up. The district celebrates. Then somebody actually talks to the kids and finds out they can memorize formulas but can't solve a problem in front of them. The metric improved. The thing the metric was a proxy for did not.

You want to get in shape, so you start counting steps. One night you're pacing your apartment at 11:54 p.m., chasing a clean 10,000. Your heart rate hasn't broken 100 since March and you haven't picked up a dumbbell in a month, but the ring on your watch is closed. Goodhart's Law.

Someone decides to measure engineering productivity by lines of code. Now you have developers inflating boilerplate, splitting one-liners into ten, and adding comments nobody will ever read. The codebase grows. The product does not improve. The original sin is the metric.

You reward sales reps for number of calls. Now everyone's voicemail is full of robotic greetings. No one is actually having a conversation. The dashboard is beautiful.

The pattern is always the same. Numbers feel objective. They give the people running the system a sense of control in a messy world. But the moment people figure out which numbers matter, they adapt. They find shortcuts. They route around the cracks. The map becomes the territory. And then the territory falls apart.

The smarter and more resourceful the people in the system, the faster the metric collapses into self-parody. This is not a moral failing. It's the system working as designed.

What to do instead

You can't escape Goodhart entirely. You can build systems that fail more gracefully when it strikes.

Use more than one metric. A single number is a plane with one dial. Track complements. If you measure speed, also measure quality. If you measure engagement, also measure retention. The point isn't dashboard bloat. The point is that any one number is easier to game than two numbers in tension.

Measure outcomes, not proxies. People game proxies because proxies are easier to manipulate than the thing you actually care about. Lines of code is a proxy. Working features in production is closer. User adoption and retention is closer still. Each step toward the real outcome makes the metric harder to fake and harder to game.

Refresh metrics on a schedule. Every metric decays as a signal once people start optimizing for it. Treat them like perishable goods. Review and rotate. The half-life of a useful KPI is shorter than people think.

Talk to humans. Surveys, interviews, slack threads, postmortems. People will often tell you exactly how they're gaming the system if you bother to ask. They're frequently relieved someone finally noticed.

Build feedback loops where misalignment is loud. Anomaly detection. Dashboards that flag surprising context. Easy escalation paths when something feels wrong. The goal is to make the gap between the metric and reality visible before it becomes structural.

Reward intent, not just output. Curiosity, learning, iteration. The strongest cultures don't let you win by optimizing a dumb metric. The reward function is messier and slower and more subjective, and that's part of why it's harder to break.

Assume someone is gaming the gaming. Eventually somebody will optimize how metrics are chosen, or who reports them, or what they include. Preempt that by being transparent about how metrics are constructed and by rotating who defines targets.

The actual lesson

Goodhart's Law is a warning that the legibility of a system and the health of a system are different things. The dashboard is not the company. The test score is not the education. The step count is not the workout. When you confuse them, you optimize for the legible version and quietly lose the rest.

The fix is not to stop measuring. The fix is to remember that you're measuring at all, and to stay loyal to the messy thing the measurement was supposed to represent.