Goodhart's Law: When the Measure Becomes the Target
observation · Goodhart's original observations in monetary policy (1981)
When a metric is used as a target for optimization, it ceases to be a good measure of the thing it was intended to measure. Optimizing for the measure rather than the underlying goal creates perverse incentives and system failure.
Core Concepts
The Problem
Incentive systems, metrics, and measurement frameworks are ubiquitous in business and policy. Leaders use metrics to align behavior and track progress. But metrics are always imperfect proxies for what actually matters. When people are rewarded for improving the metric, they optimize for the metric regardless of its relationship to the underlying goal.
The Claim
Any metric that becomes the explicit target of optimization will eventually become a poor measure of the thing it was intended to measure, because people will find ways to game the metric that don't improve the underlying outcome.
Key Evidence
- •Soviet manufacturing quotas measured in weight, not quality — factories produced useless heavy goods
- •Healthcare facilities measured on speed-to-diagnosis, not health outcomes — patients rushed through with incorrect diagnoses
- •Social media platforms optimized for engagement time, not user wellbeing — algorithmic amplification of divisive content
- •Bittensor subnet 54 miners incentivized for successful identity attacks, not identity system security — attacks increase, undermining the network's core function
Practical Implication
When designing incentive systems (compensation, metrics, KPIs), the measure you choose will drive behavior more than the stated goal. The system will optimize for the measure, not the objective. This requires: (1) choosing metrics as proxies with explicit understanding of their limitations, (2) rotating metrics frequently so they can't be gamed, (3) maintaining diverse measures of success rather than single-metric optimization.
Nuance & Limits
Not all metrics are bad — they're necessary for coordination at scale. The problem is mono-optimization: relying on a single metric as the target. Robust systems use multiple, conflicting metrics (profitability AND customer satisfaction AND employee wellbeing) so gaming one automatically creates pressure elsewhere.
Source Material
Citation Density
Very high — appears across economics, management, public policy, technology
Gaps
- ⚠ Limited exploration of how to design metric systems that are robust to gaming
- ⚠ Underexplored: positive examples of metric systems that successfully maintained alignment between measure and goal over time
Citation Trend
Who's Talking About This
15 episodes reference this idea.
Public or internal rankings of token efficiency can push teams toward penny-wise, pound-foolish optimization.
A single AI token used for casual brainstorming creates vastly different economic value than a token used to diagnose disease or optimize supply chains, yet a flat token tax would treat both identically.
Google's search has worsened not because of poor engineering but because the company's business model (ad revenue) fundamentally conflicts with its product goal (helpful search results).
The hosts discuss 'metricmaxxing'—an obsessive focus on optimizing measurable metrics at the expense of deeper goals.
Systems that optimize relentlessly for measurable metrics inevitably destroy the unmeasurable human values that make those systems worth preserving — trust, resilience, meaning, and craft.
Beijing's willingness to abandon job creation targets reveals how closely economic metrics are tied to state legitimacy—and how dangerous it becomes when the government can no longer hit those targets reliably.
Discuss Further
Open this concept in an AI assistant for deeper discussion, critique, or exploration.