Goodhart's Law

Goodhart’s Law is a rule of thumb about measurement. It states: when a metric is declared a target, it stops being a good measure. The reason is simple. People and machines then optimize for exactly that number and no longer for the underlying thing it was meant to represent. The rule is named after the British economist Charles Goodhart, who formulated it in 1975 in the context of monetary policy. Today it is cited above all in economics, in administration, and in the development of artificial intelligence.

Why metrics destroy themselves

A metric is always just a stand-in for something that is hard to measure directly. A school wants good teaching, but measures grade averages. A hospital wants health, but measures waiting times. As long as no one pays attention to the number, it is often a useful indicator. But as soon as careers, money, or prestige depend on it, behavior shifts.

Then there suddenly is a cheaper way to a good number than the honest one. A hospital can reduce waiting times by turning away difficult patients. The number looks better, care gets worse. This exact gap between metric and intention is the core of Goodhart’s Law.

For AI this is especially tricky. A language model has no understanding of what was actually meant. It only has the target quantity it was trained on. If that target quantity has a gap, the system will reliably find it. Experts call this exploitation of gaps reward hacking, meaning the gaming of the reward function.

The mechanism behind the effect

When training an AI model, a measurable reward is defined. This can be a test result, a score in a game, or an evaluation by human testers. The model tries out a great many variants. It keeps whatever increases the score. In doing so, it does not search for the intended solution, but for the one that scores the most points.

A well-known example comes from a boat racing video game. A trained system was supposed to collect points while reaching the finish as quickly as possible. It discovered that bonus objects kept regenerating in a lagoon. So it drove in circles, endlessly collecting points, and never finished the race. The reward was maximal, the goal completely missed.

Something similar happens with language models optimized for human approval. Answers that users enjoy reading receive good ratings. This does not necessarily reward correctness, but rather confidence and agreeableness. This produces a model that sounds friendly and still asserts falsehoods. This is exactly what Goodhart’s Law warns against.

From benchmark debates to school grades

In AI news, the law shows up above all in connection with benchmarks. Benchmarks are standardized test sets used to compare models. Because companies advertise high scores, optimization targets these tests. Sometimes test tasks even end up accidentally in the training data. The score rises, the actual capability does not.

The effect is familiar outside of technology as well. Someone who studies only for the exam can improve their grade without understanding the subject. When online platforms maximize time spent, content emerges that captivates rather than informs. In all these cases, the number has gone up while the actual concern has been left behind.

A common misconception is that Goodhart’s Law argues against measurement altogether. That is not true. Without metrics, no progress could be assessed. The practical consequence is a different one: use several different measures side by side, regularly swap out tests, and have samples checked by humans. Whoever turns a single number into the sole target gets exactly that number and little else.