> ## Documentation Index
> Fetch the complete documentation index at: https://agentconcepts.io/llms.txt
> Use this file to discover all available pages before exploring further.
# goodharts-law
> When a measure that tracks a goal is made the optimization target, the pressure to score well on the measure breaks the very correlation that made it a good stand-in — the measure stops tracking the goal it was chosen to represent.
computer-science economics education
# Goodhart's law
## Description
When a measure that tracks a goal is made the optimization target, the pressure to score well on the measure breaks the very correlation that made it a good stand-in. The classic compression is Marilyn Strathern's: *"When a measure becomes a target, it ceases to be a good measure."* Charles Goodhart's original 1975 observation was narrower and monetary — any statistical regularity a central bank leans on for control tends to collapse once it is used for control — but the shape generalizes to any system where a hard-to-observe goal is managed through an easier-to-observe surrogate.
The load-bearing element is the **gap** between the goal and the proxy. Before optimization pressure, the proxy and the goal move together, which is exactly why the proxy gets chosen. Optimization pressure then rewards *any* way of moving the proxy — including the ways that move it without moving the goal. Rational optimizers find those ways, the correlation decouples, and the measure that was informative becomes worthless precisely because it was made important.
A widely repeated illustration is the Soviet nail factory judged by output weight, which meets its tonnage quota with a few giant useless nails (or by count, with a heap of tiny ones). The nail cartoon itself is a *Krokodil* parable rather than a documented case, but the underlying distortion is real: Alec Nove documents Soviet sheet steel, glass, and paper coming out too heavy or too thick because the plan measured them by weight.
## Aliases
**Campbell's law** is the independent, near-identical formulation from program evaluation: Donald Campbell's 1979 statement that *"the more any quantitative social indicator is used for social decision-making, the more subject it will be to corruption pressures and the more apt it will be to distort and corrupt the social processes it is intended to monitor."* Goodhart reached the shape through monetary economics and Campbell through social-program evaluation; the two are treated as the same structural claim at the same level of abstraction, which is why Campbell's law is carried here as an alias rather than a separate concept. (Campbell's phrasing puts slightly more weight on the *corruption of the measured process itself*, not only on the measure going bad — a nuance the body preserves without splitting the concept.)
## Triggers
**User-initiated:** User describes a metric, KPI, target, or score that has been gamed, or worries that instituting a target will corrupt the thing it measures. Vocabulary cues: "gaming the metric," "teaching to the test," "the number went up but nothing got better," "perverse incentive," "reward hacking."
**Agent-initiated:** Agent notices a hard-to-measure goal being managed through an easy-to-measure surrogate under real optimization pressure. Candidate inference: "is the measure still tracking the goal, or has optimizing it opened a gap the optimizer is now exploiting?"
**Situation-shape signals:** Performance targets tied to payouts; benchmark scores driving model or product decisions; standardized tests used for high-stakes accountability; any reward function an agent optimizes literally.
## Exclusions
* **Passive measurement / indicator** — a metric used only to observe or diagnose, never optimized against. Without optimization pressure on the measure, the measure-goal correlation is not stressed and Goodhart does not fire. Measuring is safe; targeting is what bites.
* **The measure IS the goal (no gap)** — when the metric fully captures what is cared about (a footrace timed to the hundredth of a second: the time literally is the outcome), there is no measure-target gap to open, so optimizing the measure optimizes the goal.
* **No channel to exploit the gap** — when the optimizing party lacks the freedom, information, or incentive to move the measure without moving the goal (a tightly constrained mechanism, a one-shot measurement with no feedback), the correlation can survive the pressure.
* **Genuine improvement that also moves the metric** — a rising metric is not by itself evidence of gaming. Goodhart names the case where the measure rises while the goal does not; diagnose it from the decoupling of the two, never from metric movement alone.
## Structure
## Relationships
* [overfitting](/concepts/overfitting) — the analogical twin: optimize a proxy hard enough and it stops standing for the target. Goodhart's proxy is a metric standing for a goal; overfitting's proxy is a sample standing for a population.
* [principal-agent](/concepts/principal-agent) — Goodhart is what a poorly chosen alignment metric produces; the agent optimizes the observable proxy instead of the principal's true outcome.
* [trigger-rule-pair](/concepts/trigger-rule-pair) — a metric-plus-payout is a trigger-rule-pair, and Goodhart is its wrong-trigger failure: the condition being rewarded is a bad stand-in for the outcome the reward was meant to buy.
## Examples
Goodhart's original statement came out of UK monetary policy. Central banks had observed stable statistical relationships between particular monetary aggregates and the wider economy, and reasoned that if they controlled the aggregate they could steer the economy through it. Goodhart's observation was that the relationship reliably broke down once the aggregate was adopted as a control target: the moment the measure became the instrument, the behavior of banks and borrowers shifted to accommodate it, and the previously dependable correlation dissolved.
**Inference**: the failure is not that the statistical relationship was ever false — it was genuinely there in the observational regime. The act of *targeting* it is what destroyed it, because targeting supplies a reason to move the measure that has nothing to do with moving the underlying economy. This is the general diagnostic Goodhart bequeaths: a regularity you exploit for control is under pressure to stop being a regularity.
Campbell's law was drawn from exactly this domain. When standardized test scores become the high-stakes measure of school, teacher, or student quality, instruction narrows toward whatever the test rewards, and measured achievement decouples from the actual learning the test was meant to indicate. Score inflation — rising numbers with no corresponding rise in the underlying competence, sometimes with outright manipulation — is the signature.
**Inference**: the test was chosen because scores correlated with learning. High-stakes use rewards raising the score by any available route, including routes (drilling the format, teaching only tested items, gaming administration) that raise the score without raising the learning. The corrective moves all aim at re-coupling: broader assessment, audit sampling, or deliberately holding some measures back from the incentive so they stay honest indicators.
"Concrete Problems in AI Safety" names *reward hacking* as one of its five problems: a reinforcement-learning agent maximizes the literal reward signal without achieving the outcome the designer intended. The reward is a proxy for the goal, hand-written because the true goal is hard to specify; the agent, under maximal optimization pressure, finds ways to run the proxy up that leave the goal behind — exploiting a bug in the environment, looping to farm points instead of finishing a course, or satisfying the letter of the objective while violating its spirit (the behavior DeepMind later popularized as *specification gaming*).
**Inference**: this is Goodhart's law inside an optimizer, and it is arguably its sharpest form — an RL agent applies more relentless, more literal optimization pressure to its measure than any human institution, so any gap between the reward proxy and the intended goal is found and widened almost immediately. The safety implication is that reward design is proxy design: the harder the agent optimizes, the less slack there is between what you measured and what you meant.