Homeglossary

Goodhart's law

Goodhart's law is a measurement principle that says a metric stops being a reliable measure when it becomes the target people are rewarded or pressured to hit. It occurs because people adapt their behavior to improve the number, sometimes without improving the real outcome the number was meant to represent. The law does not mean organizations should abandon metrics, but that they should design and review them with care.

What Is Goodhart's Law?

Goodhart's law is the principle that a metric stops being a reliable measure when people are rewarded or pressured to hit it as a target. People may improve the number without improving the real result it was meant to represent.

A metric is a measurement, such as response time, sales, test scores, or website visits. A target gives it consequences, such as a bonus, ranking, funding decision, or performance review. Metrics are usually proxies, or imperfect stand-ins, for harder-to-measure goals such as trust, learning, health, or quality. Goodhart's law does not say measurement is useless. It says metrics need context, guardrails, and review.

Origin and Meaning of the Famous Quote

British economist Charles Goodhart introduced the underlying idea in 1975 while discussing monetary policy. His observation was that a statistical pattern can break down once authorities use it to control a system. People and institutions react to the rule, changing the behavior that made the pattern useful.

The familiar wording, "When a measure becomes a target, it ceases to be a good measure," is widely associated with anthropologist Marilyn Strathern. It is useful shorthand, not an absolute rule. A measure can remain helpful as part of a broader assessment. The danger comes when decision-makers treat it as the whole goal.

How Goodhart's Law Works

The pattern often develops gradually. It can result from deliberate manipulation, but also occurs when well-intentioned people focus on what the system rewards.

  1. Define the real objective, such as resolving customer problems or improving student understanding.
  2. Choose a proxy metric related to that objective, such as average handling time or exam pass rate.
  3. Attach pressure, rewards, or penalties to the metric.
  4. People adapt their work to improve the reported number, intentionally or unintentionally.
  5. The metric rises while the original objective stagnates or worsens, so the proxy no longer gives a trustworthy picture.

The Four Types of Goodhart's Law

Researchers often describe four forms of the problem. The categories can overlap, but they help explain why a metric failed.

TypeFailure mechanismExample
RegressiveA proxy is correlated with the goal, but imperfectly. Selecting the highest scores also selects noise.Hiring only applicants with the highest test scores misses candidates with strong practical skills.
ExtremalThe relationship between metric and goal stops holding at extreme values created by optimization.Publishing more articles may help reach at first, but excessive output can reduce editorial quality.
CausalChanging the measure directly does not create the outcome that the measure once predicted.Giving people a certificate may signal expertise, but issuing more certificates does not create expertise.
AdversarialSomeone knowingly exploits the rules or measurement method to score well.A seller generates fake reviews to raise a marketplace rating.

Goodhart's Law Examples in Everyday Systems

Goodhart's law appears wherever a simple score stands in for a complex outcome. The key question is: what does the target miss?

  • Business KPIs: A call center rewards short calls. Agents may end calls quickly and leave customers without help.

  • Sales: A team is paid for signed contracts. Representatives may discount heavily or sell unsuitable plans instead of creating sustainable revenue.

  • Education: A school is judged mainly by test scores. Teachers may narrow instruction to test content rather than build broader understanding.

  • Healthcare: A clinic targets shorter waiting times. Staff may improve the recorded wait by changing check-in rules rather than increasing access to care.

  • Public policy: A program rewards job placements. Short-lived placements can inflate the count without creating stable employment.

  • Search and social media: Creators optimize clicks or watch time. Sensational material can win attention more easily than useful, trustworthy information.

  • Software delivery: A team is judged by tickets closed or code shipped. Quick output can create defects and maintenance work. Teams building with an AI app builder still need to assess usability, security, and reliability, not just release speed.

  • AI evaluation: A model is optimized for a benchmark score. It may learn benchmark-specific patterns while remaining unreliable on unfamiliar real-world tasks.

Goodhart's Law vs. Campbell's Law and the Cobra Effect

These concepts overlap because all warn about harmful incentives. Their emphasis differs.

ConceptMain concernExample
Goodhart's lawA proxy loses value when it becomes a target.Closing tickets becomes more important than solving problems.
Campbell's lawHeavy reliance on a quantitative indicator invites corruption and distorts the process it monitors.High-stakes testing changes teaching and can undermine learning.
Cobra effectAn incentive creates an unintended outcome that makes the original problem worse.A bounty for pests encourages people to breed pests for payment.
Perverse incentiveAny reward or rule that motivates behavior contrary to the intended goal.A bonus for volume encourages low-quality sales.

Benefits of Using Metrics Carefully

Metrics remain essential because organizations cannot manage complex work through intuition alone. Used as evidence rather than unquestionable truth, they offer real value.

  • They focus attention on important outcomes and make priorities visible.
  • They reveal trends and operational problems earlier than anecdotes alone.
  • They support accountability when leaders also examine quality and context.
  • They enable learning by showing whether a change likely improved results.
  • They help teams compare alternatives, especially when a well-designed AI dashboard builder presents related signals instead of one headline number.

Limitations and Common Pitfalls

A dashboard can look precise while omitting the outcome people care about. Adding metrics helps only if they expose different risks rather than repeat the same blind spot.

  • Target fixation makes people pursue the number after it no longer serves the mission.
  • One-number management hides trade-offs, such as speed versus accuracy or sales versus retention.
  • Easy-to-count activity can crowd out difficult-to-measure quality, trust, and long-term effects.
  • Lagging harms may remain invisible until after rewards are paid or decisions made.
  • Comparisons can be unfair when teams serve different customers, regions, or levels of difficulty.
  • Manipulation can range from falsifying data to changing workflows solely to improve reporting.

How to Avoid Goodhart's Law When Setting KPIs

No metric design eliminates every unintended response. A practical governance process reduces the chance that a target will distort the work.

  1. State the real outcome in plain language before choosing a KPI.
  2. Select a balanced set of metrics covering results, quality, cost, and risk.
  3. Pair leading indicators, which provide early signals, with lagging indicators that show final results.
  4. Add guardrails, such as customer satisfaction, error rates, complaints, or safety checks.
  5. Review real cases behind the numbers, including unusually high-performing cases.
  6. Audit for gaming by asking how someone could improve the score without delivering value.
  7. Refresh targets when conditions, incentives, or customer needs change.
  8. Give employees a safe way to report when a metric causes harmful behavior.

A Practical Test for a Healthy Metric

Before adopting a target, apply a five-part test: relevance, resistance, coverage, review, and trust. Ask whether the metric relates to the outcome, is difficult to manipulate, captures likely harms, has a review date, and would still be trusted if leaders knew exactly how the score was achieved.

If the final answer is no, the metric needs a companion measure or redesign. Goodhart's law is not an argument against KPIs. It is a reminder that metrics should inform human judgment, not replace it.

Start Building
on Emergent today
Start Building