Measuring Goodhart’s law

OpenAI’s blog post explores the implications of Goodhart’s law for AI development. The law, originally from economics, states that when a metric becomes a target, it loses its validity as a measure. OpenAI acknowledges that this is a central challenge when optimizing objectives that are expensive or difficult to quantify directly. For instance, using proxy metrics like user engagement to represent value can lead to gaming the system or unintended behaviors. The post reflects on how such pitfalls arise in reinforcement learning from human feedback (RLHF) and other training paradigms. For developers building AI workflows, this serves as a reminder to carefully select evaluation metrics, monitor for reward hacking, and maintain alignment between proxy measures and true goals. Rather than offering a solution, OpenAI frames the issue as an ongoing tension that requires constant vigilance and iteration.