When the metric becomes the goal

The world is full of systems that optimize something. Schools optimize grades, companies optimize profit, governments optimize growth, and platforms optimize engagement.

What interests me is that these systems often end up optimizing the wrong things, or optimizing simplified metrics that stand in for richer goals.

What should a student optimize for? A natural answer is understanding. But in practice, the student optimizes for passing exams or getting good grades.

That is understandable. Grades are clear, legible, and comparable. They make it easy for schools to evaluate students, and they make it easy for students to compare themselves to one another.

But that does not mean they reflect what we actually value properly. A student can get high grades while understanding very little, forgetting everything after the exam, or losing curiosity in the process.

There are two layers to the problem.

The first question is what we choose as the goal itself. If we define the goal as nothing but maximizing grades, we have already chosen a narrow objective. If instead we choose real understanding, long-term retention, curiosity, and the ability to think independently as the goal, we choose something richer and more aligned with what education should actually care about.

Then comes the second problem. Those richer goals are hard to measure directly, so schools and students often rely on proxy metrics such as grades that are meant to stand in for actual understanding.

At first, this works well. But once systems start optimizing directly for the proxy, the metric can drift away from the thing it was meant to represent.

If a student optimizes for grades, they might focus on memorization, test-taking strategies, or even cheating, rather than on deep understanding. The system becomes good at maximizing the metric while becoming worse at achieving the real goal.

That sets up the central dynamic of the article: a system can become very good at maximizing a number while becoming worse at achieving the thing that number was supposed to represent.

Why are we optimizing for the wrong things?

This dynamic has a name: Goodhart’s Law. It states that:

“When a measure becomes a target, it ceases to be a good measure.”

Initially, proxy metrics often correlate with the real goal. But once we start optimizing directly for them, behavior adapts.

At first, grades may correlate reasonably well with learning. But under strong optimization pressure, the study strategy that produces the best grades is no longer necessarily the one that produces the deepest understanding.

A study strategy can look successful because it increases grades, while at the same time producing shallow understanding, short-term memorization, or intellectual passivity. None of that necessarily appears in the metric being optimized. The number goes up, but the wider outcome may get worse.

What should we optimize for instead?

The problem is not that systems optimize. Optimization is powerful and necessary. The real problems are that we can choose a metric as the objective itself, and then optimize a metric that is not a proper representation of the real goal.

Grades are not inherently a bad metric. But when students optimize for grades instead of understanding, the system becomes more about maximizing a number than about learning. The same pattern appears elsewhere too. Companies optimize profit instead of the actual value they create for society. Social media platforms optimize engagement instead of meaningful interaction or truthful information. Governments optimize GDP (Gross Domestic Product) growth instead of broader prosperity. In each case, either the objective is already too narrow, or the metric starts replacing the richer goal it was supposed to represent.

So the solution is not to remove optimization. The solution is first to ask whether we are pursuing the right goal at all, and then to design better ways of measuring progress toward it.

In education, that would mean caring less about grades alone and more about whether students actually understand, retain, question, and apply what they learn.

The point is not that any single measure will be perfect. The point is that if we know an objective is too narrow, we should revise it. And if we know a metric is incomplete, we should try to find a more comprehensive representation of the objective. We should build systems that reflect more of what we actually value.

The difficulty, of course, is to define metrics that approximate our goals well enough that, when a system optimizes for them, it produces more of what we actually want rather than more unintended harm.

That is not easy. But if we want better systems, then we need better ways of quantifying the concepts.

In the end, systems will always optimize something. The question is whether we are deliberate enough to choose objectives worth optimizing, and careful enough to measure them in ways that do not strip away what matters most.