Goodhart Law Gaming

Once a goal becomes the rule for judging success, people change their behavior to meet that rule. That shift is exactly what makes the rule a poor guide to the real goal.

Goodhart-style gaming occurs when an operational metric becomes the target itself, inducing agents to optimize for the metric rather than the underlying objective. It leads to metric capture, where the proxy diverges from true performance because of adaptive responses and incentive distortion.

A school district measures teacher quality by student test scores, so teachers begin spending all instructional time on test-prep drills. Scores rise, but students' actual understanding of the subject weakens — the metric improved while the real goal was undermined.

A hospital system adopts 30-day readmission rate as its primary quality KPI under a value-based purchasing contract. Clinical teams respond by extending initial admissions beyond medical necessity, discharging patients to skilled nursing facilities that reset the 30-day clock, and selectively avoiding high-risk patient panels whose comorbidities inflate readmission probability. The proxy metric improves on paper — satisfying the payer's compliance algorithm — while true care continuity and population health outcomes degrade. It's a textbook case of metric capture: the proxy's centrality in the incentive architecture channels agent strategies entirely toward optimizing the proxy, producing a systematic divergence between the observed KPI values and the latent objective of durable patient health.

Once the score is used as the target, people change their actions to boost that score. Those actions are exactly what makes the score stop matching the real goal.

The proxy metric sits within the observation channel and gets selectively emphasized, creating asymmetrical incentives that weight actions by their effect on the proxy. That structural weighting yields adaptive behavior and constraint exploitation, producing a divergence between the observed proxy values and the latent objective.

Change how success gets measured so it captures more of the real goal. Watching for new tricks and adjusting the measure over time keeps it honest.

Composite metrics and randomized audits reduce single-proxy domination and help detect manipulation. Adaptive evaluation that reweights signals and penalizes narrow optimization strategies closes the gap further.

proxy_capture; perverse_incentives; gaming_externalities

An adversarial actor can deliberately design or lobby for a proxy metric they know is easy to game, then exploit the resulting divergence between the metric and the true objective for private gain while appearing compliant—e.g., a contractor hitting numerical delivery targets while degrading actual service quality. In competitive or regulatory contexts, sophisticated actors can reverse-engineer evaluation rubrics to surface-optimize for inspectable signals, effectively weaponizing auditors' reliance on the proxy against the auditing institution itself. This is especially potent in opaque environments where the principal cannot cheaply observe the latent objective, enabling sustained metric capture with minimal detection risk.

Rotate and randomize the composition of composite metrics over time to prevent agents from locking in on stable gaming strategies, and supplement quantitative proxies with qualitative audits that assess the latent objective directly. Build adaptive evaluation systems that penalize sudden, unexplained improvements in a narrow proxy (a statistical flag for gaming) and reward consistency across multiple independent signals. Structural separation of those who define metrics from those who benefit from them reduces principal-agent conflicts that enable proxy capture.