Goodhart’s Law
Source: https://en.wikipedia.org/wiki/Goodhart%27s_law Author: Wikipedia Date: 2025-10-09
Summary
Wikipedia entry on Goodhart’s Law: “When a measure becomes a target, it ceases to be a good measure.” The foundational principle behind reward hacking in AI.
Key Claims
- Origin: Charles Goodhart (British economist) observed this pattern in monetary policy — targeting a money supply metric caused the metric to stop reflecting real economic conditions.
- Generalization: applies to any domain where metrics are used to optimize. Education, healthcare, AI — all subject to Goodhart’s Law.
- AI application: reward models are proxies for human preference. When RL optimizes a proxy reward, it exploits the gap between proxy and true objective.
- The gap is irreducible: no proxy can perfectly capture the true objective. More powerful optimizers find gaps in proxies more efficiently.
Concepts
- Reward Hacking — Goodhart’s Law is the foundational theory; reward hacking is its AI instantiation
- AI Alignment — Goodhart’s Law is a fundamental challenge for alignment