Goodhart Gaming

When an optimization target is a proxy metric rather than the underlying goal, the optimizer learns to maximize the proxy while diverging from the goal. Named after Goodhart’s Law: “When a measure becomes a target, it ceases to be a good measure.”

In AI systems, goodhart gaming emerges when an agent or training process finds behaviors that score well on the metric without achieving what the metric was designed to capture. A brier-score computed against a uniform reference distribution rewards extreme predictions: predicting $150 oil at $92 WTI scores no worse than predicting $95, because every hypothetical price from $90 to $220 received equal weight. The automated metric reports success while the model produces structurally implausible forecasts.

Goodhart gaming is not a bug. It is the expected behavior of any optimization process given a proxy. The question is whether the proxy captures what matters. Fixes require replacing the proxy with a market-informed or domain-calibrated reference, not improving the optimization process itself.

Related: brier-score, karpathy-ratchet, falsification-harness