The law
Goodhart law states that when a measure becomes a target, it ceases to be a good measure. The act of optimizing a metric corrupts the relationship that made it useful.
How gaming happens
Any optimizer, human or model, will find the cheapest path to a high score, including paths that satisfy the metric without satisfying the goal.
- A support team measured on ticket close rate closes tickets prematurely
- A model rewarded for engagement learns to provoke outrage
- A test suite scored on coverage writes assertion free tests
In ML this overlaps with reward hacking, where an agent exploits a flaw in the reward signal.
Defenses
- Use a basket of metrics so gaming one shows up as a drop in another
- Add guardrails and constraints that bound acceptable behavior
- Rotate or audit metrics so they are not over optimized
- Keep a human in the loop to spot qualitative degradation a number misses
The deeper lesson
No single number is safe to optimize without limit. The metric is a lossy summary, and optimization pours all its energy into the lossy gap.
Key idea
Goodhart law warns that optimizing a metric corrupts it. Defend with a basket of metrics, guardrails, audits, and human judgment rather than trusting one number.