# What does Godhart’s law mean?

284 views

It goes “When a measure becomes a target, it ceases to be a good measure.” How does that work in practice?

In: Other

In practice, it means that when a particular metric or measurement is used as the sole target or goal, people may focus solely on achieving that target, often at the expense of other important factors or the overall purpose of the measure.

For example, in a business setting, if profit margin is set as the sole target, employees may prioritize short-term gains or cut corners to meet that target, even if it harms long-term sustainability or customer satisfaction.

In essence, when a measure becomes the primary target, people may start optimizing for that measure without considering broader implications or the original intent of the measure. This can lead to distorted priorities and unintended consequences.

When we perform a task, we would also like to know how well we perform that task. For example, almost everyone has gone to school. In school, you take a class. In the class, you learn something. At the end of that class, we measure how well you learned the something, and call it the grade. We use the grade to determine other things about you as well – if you have good grades, then you can go to a good college, for example.

Well, now the grade is the target, not the learning. So everyone is interested in getting the best grade, as opposed to learning the most. So rather than working harder, people find a way to get a better grade that doesn’t involve learning. They cheat with ChatGPT. The argue with their teacher / professor about their grade. They complain to the administration of the school / college.

Once, an “A” meant you did really well, a “B” was average, and a “C” was satisfactory. Now, an A is about average in many situations, and anything less seems unsatisfactory. Grades no longer measure how well you learned something – just about everyone has an A. The grade is the target, and is is no longer a good measure of the learning.

Goodhart’s Law is like saying if you’re playing a game and the score is just for fun, you play differently than if the score decides who wins a prize. When the score becomes the way to win, you might play just to get points, not to enjoy the game or play well.
In real life, if a school decides to judge teachers by their students’ test scores, teachers might just teach to the test. The scores go up, but it doesn’t mean kids are learning better overall. The test score stops being a good sign of real learning because it’s now a target, not just a measure

So, Goodhart’s Law warns us that when we turn a measurement into a goal, it can stop showing what we originally wanted to measure because people start changing their behavior to meet the goal, not to improve the actual thing we care about

I worked in customer support back in the day.

When they were solely focused on average handle time, and tied performance only to average handle time, that lead to folks with complex issues that would take time to investigate to be punted around between different agents that didn’t want their time measurement and therefore their money affected. Because the count was solely based on how little time you spent on tickets, not on your quality of response, we had a problem with folks constantly giving people wrong information because they didn’t want to spend the time to find the right answer. Our customer satisfaction rate plummeted, we lost customers, and we had people waiting months to get an answer because nobody wanted to spend the time to do it, because they were rewarded for how little time they spent.

For example, suppose I run a company whose job is to clean up park. I get the idea of rewarding workers per square meter they cleaned. This can have the unintended consequence of workers rushing through the park with their brooms, maximizing the area while doing poor job.

Basically, if you set a metric that doesn’t absolutely perfectly capture your goal, you run a risk of people trying to game the system and maximize that metric alone while neglecting other important aspects of their jobs.

Imagine that you notice that all your customers that are happy have talked to a sales associate in the last thirty days. Then you check to see which sales associates talk to clients more often. You notice the ones that talk more also sell more. That’s a good predictive measure of who is trying the hardest. Then imagine you tell everyone that if they haven’t talked to their clients, in the last three days they get fired. Sales associates will make a bunch of annoying and useless calls, and it’s likely worse for you than before, and now your metric doesn’t mean anything because you can “game the system”.

Here is one of my favorite examples: in the early days of covid, while everybody was trying to figure out what to do with public spaces, schools, etc., one school found out that the official definition of “close contact” (for whoever was running their show at the time) was “within 6 feet of someone for 15 minutes.”

Their solution to avoiding close contact in the classroom was to make all the children change seats every 12 minutes. Thus, no close contact at all, though of course it would only make viral transmission worse.

This is the heart of Godhart’s law: you see a problem, make up a rule or metric to try to address it, and very often the target audience will find a way to follow the rule or optimize the metric in a way that doesn’t solve or even worsens the original problem.

Here’s a very real example from when I managed a McDonalds.

The staff are scored on how quickly they serve food.

An order comes in, and is displayed on a monitor. The food is made, then pushed out to the front counter. once the order leaves the kitchen they “clear” the order, removing it from the monitor.

Should work just fine.

In practice the first person looks at the screen, quickly memorizes what’s on it, and then clears the order. Then they start preparing the food.

The time creeps up a bit as more orders come in and the staff don’t immediately clear them. They do get in trouble for screwed up orders after all. They might get the order started by laying out wrappers, and putting special request tickets down to help them “remember” the order before clearing it. But the order is getting cleared long before it’s actually finished.

So the order time is no longer a valid measure of the performance of the crew. It’s the goal, and no longer measures the time to make the food.

I’d get a lot of push back from general managers because my times where “high” as I worked hard to keep the crew on my shifts from clearing the order early, but I pointed out my were honest and useful. And we rarely had huge crisis in the kitchen from botched orders etc.

I’ve experienced this multiple times in my working career.

If you measure the things that people do, then incent people to hit particular values (or, even worse, penalise them for not hitting them), you are also incenting them to game the system in order to succeed, and should expect them to do so. At which point the measurements are no longer useful, because they aren’t trustworthy.