← Glossary Term
Goodhart's law
When a measure becomes a target, it stops being a good measure.
The economist Charles Goodhart observed that as soon as people are rewarded for a number, they optimise the number rather than the thing it was supposed to represent. Measure lines of code and you get verbose code. Measure tokens consumed and you get people running AI on tasks that did not need it.
It comes up constantly in AI, in two places. Inside training, a model rewarded on a proxy for helpfulness learns to look helpful. Outside it, organisations that track AI adoption instead of AI results get adoption. Meta’s retreat from counting tokens in performance reviews in September 2026 is a clean example of the law catching up with a metric.
Mentioned in