Goodhart's Law

When a measure becomes a target, it stops being a good measure.

A metric is useful only while it is being used to observe reality. Once people begin optimizing the metric itself, they often discover ways to increase the number without improving the thing the number was originally meant to measure.

Metrics are proxies for reality

• Test scores are a proxy for learning.
• GitHub commits are a proxy for engineering productivity.
• Citation counts are a proxy for scientific impact.
• GDP is a proxy for economic well-being.

As long as nobody is directly optimizing these numbers, they often correlate well with what we actually care about. Once rewards, promotions, or status become tied to them, people naturally optimize the proxy instead of the objective.

Common examples

Education

• Goal:
  • Students genuinely understand the material.
• Metric:
  • Standardized test scores.
• Once schools are evaluated primarily on scores:
  • Teachers teach only what appears on the exam.
  • Students memorize instead of understanding.
  • Difficult to test or creative topics receive less attention.

Software Engineering

• Goal:
  • Build useful, maintainable software.
• Metric:
  • Lines of code written.
• Drift:
  • Developers can increase the metric simply by writing more code.
  • The codebase becomes larger. Productivity becomes lower.

Research

• Goal:
  • Discover new knowledge.
• Metric:
  • Number of published papers.
• Researchers may:
  • Split one project into several smaller papers.
  • Choose fashionable topics over important ones.
  • Optimize for publication frequency.

Social Media

• Goal:
  • Inform or entertain people.
• Metric:
  • Clicks and engagement.
• Drift:
  • Creators learn that outrage, controversy, and sensational headlines attract more attention.
  • Engagement increases. Content quality decreases.

AI Benchmarks

• Goal:
  • Build generally capable AI systems.
• Metric:
  • Accuracy on a benchmark such as MMLU, or SWE-bench.
• Drift:
  • Researchers spend increasing effort optimizing specifically for benchmark performance.
  • Eventually, models become exceptionally good at those particular datasets while cannot generalize on real-world tasks.

Why this happens

Initially, the relationship looks like this:

Underlying Goal
       ↑
     Metric

The metric reflects reality.

Once incentives are attached:

Underlying Goal
Metric ← directly optimized

People optimize what is measured rather than what is intended. The correlation weakens. Eventually, the metric measures its own optimization instead of the original objective.

Goodhart's Law is fundamentally about proxy optimization

Every measurement is an imperfect representation of reality. Once optimization pressure becomes sufficiently strong, people begin exploiting imperfections in the proxy rather than improving the underlying objective.

This is why every high-stakes metric eventually becomes vulnerable to gaming.

How to reduce Goodhart's Law

• Optimize the underlying objective, not just the metric.
• Use multiple independent metrics instead of a single number.
• Periodically update or replace metrics before they become heavily gamed.
• Combine quantitative measurements with qualitative human judgment.
• Whenever possible, replace prediction with direct observation of outcomes.