Short Definition
Goodhart’s Law states that when a metric becomes a target, it ceases to be a good measure—an effect that is amplified in machine learning systems.
Definition
In the context of machine learning, Goodhart’s Law describes the phenomenon where optimizing a chosen metric causes the metric to lose its ability to reflect the true objective it was intended to measure. This occurs because models, training procedures, and human incentives adapt specifically to the metric rather than to the underlying goal.
Optimizing the measure distorts the meaning of the measure.
Why It Matters
Modern ML systems rely heavily on metrics to guide training, evaluation, deployment, and governance. When these metrics become optimization targets—especially proxy metrics—the system may improve numerically while degrading in real-world usefulness, safety, or fairness.
Metric optimization can diverge from outcome optimization.
How Goodhart’s Law Manifests in ML
Goodhart’s Law appears when:
- proxy metrics are treated as objectives
- benchmarks become leaderboards
- thresholds are tuned solely to improve scores
- feedback loops reinforce metric-specific behavior
- evaluation ignores downstream consequences
The model learns the metric, not the task.
Types of Goodhart Effects in ML
Proxy Goodhart
Occurs when a proxy metric no longer tracks the true objective.
- example: click-through rate replacing user satisfaction
Metric Gaming
Occurs when systems exploit weaknesses in metric definitions.
- example: confidence inflation to boost calibration scores
Evaluation Overfitting
Occurs when repeated optimization targets a fixed test or benchmark.
- example: leaderboard overfitting
Feedback-Induced Goodhart
Occurs when model decisions change the data-generating process.
- example: recommender systems shaping user behavior
Metrics reshape reality.
Minimal Conceptual Illustration
True Objective → Proxy Metric → Optimization → Divergence
Relationship to Proxy Metrics
All proxy metrics are vulnerable to Goodhart’s Law. The farther a proxy is from the true objective—and the longer it is optimized—the greater the risk of metric corruption.
Proxies require constant validation.
Relationship to Offline Metrics vs Business Metrics
Offline metrics often serve as proxies for business outcomes. Goodhart’s Law explains why offline improvements frequently fail to translate into real-world gains.
Offline success does not imply business success.
Relationship to Decision Cost Functions
Explicit cost functions reduce Goodhart risk by grounding optimization in real consequences rather than abstract scores. However, even cost functions can become targets if assumptions are incorrect or outdated.
Explicit does not mean immune.
Goodhart’s Law and Model Updates
Automated retraining and continuous deployment amplify Goodhart effects by repeatedly reinforcing metric-specific behaviors unless evaluation criteria are regularly audited.
Automation accelerates distortion.
Mitigation Strategies
Common strategies to reduce Goodhart effects include:
- using multiple complementary metrics
- rotating or refreshing evaluation sets
- validating metrics against long-term outcomes
- incorporating human oversight
- stress testing beyond target metrics
- aligning metrics with explicit cost functions
Metrics must be governed, not trusted blindly.
Common Pitfalls
- treating metric improvement as goal achievement
- optimizing a single metric indefinitely
- ignoring second-order effects
- failing to revisit metric definitions
- assuming metrics are objective truths
Metrics are instruments, not objectives.
Summary Characteristics
| Aspect | Goodhart’s Law in ML |
|---|---|
| Trigger | Metric becomes optimization target |
| Effect | Metric loses meaning |
| Risk | Silent performance degradation |
| Amplified by | Automation, scale, feedback |
| Mitigation | Metric governance |
Neural Network Lexicon 