OpenAI on Measuring Goodhart's Law in AI Development

OpenAI discusses the implications of Goodhart's law, which states that when a metric becomes a target, it loses its effectiveness as a measure. The company explains how this economic principle applies to AI development, particularly when optimizing objectives that are hard to quantify. They outline their approach to addressing this challenge.

WHY IT MATTERS ↘Most AI teams optimize proxy metrics — benchmark scores, reward-model outputs, human-approval rates — that degrade once made into explicit targets, so unmanaged Goodhart effects silently convert training gains into capability or safety regressions that only surface after deployment. Because regulators and enterprise buyers increasingly treat those same benchmarks as evidence of compliance or quality, the choice of proxy becomes a competitive and governance liability, not just a technical detail.