Goodhart's Law

Goodhart's Law

Overview

Goodhart's Law: when a metric controls behavior, people optimize the metric rather than the underlying goal. Formulated by economist Charles Goodhart (1975) on UK monetary policy; sharpened by Marilyn Strathern (1997): "When a measure becomes a target, it ceases to be a good measure." Four failure mechanisms (Manheim & Garrabrant 2018): Regressional, Extremal, Causal, Adversarial. Countermeasure is always multi-metric + audit + rotation.

Composes with feedback-loops, principal-agent, okr-goal-setting, survivorship-bias.

When to Use

• A KPI is being introduced or its weight is increasing in performance evaluation
• A metric is "improving" without corresponding improvement in the underlying goal
• People are visibly optimizing for a number rather than the work it was meant to track
• Algorithmic optimization is producing outcomes the designers didn't intend
• Resource allocation is driven by a single composite score or ranking
• An AI model, benchmark, or engagement metric is being optimized (or used to justify AI capex / adoption / AI-native competition) and the score is rising faster than real capability or user value

Not when: metric and goal are identical; stakes too low for gaming; metric is purely descriptive with no reward/punishment; question is which metric to use, not whether the measurement-reward system is sound.

Coaching Novices (Adaptive Front Door)

• Engine mode: user has a concrete metric or system → run The Process directly.
• Coach mode: user is unfamiliar or has no concrete case → guide step by step.

In Coach mode, respond one step at a time. Each [WAIT] is a hard stop — output only that step's question, then stop.

1. One-line: before relying on a metric to control behavior, predict how people will game it — choose the system that survives that prediction.
2. Check fit: if the metric is purely descriptive (no reward attached), Goodhart's law doesn't apply yet.
3. Elicit the specific metric and the underlying goal: what's being measured? What's the actual outcome you care about?

[WAIT — do not advance until user responds]

1. One question at a time: proxy gap? How would a clever agent game this? Which Goodhart category? What countermeasure fits?

[WAIT — do not advance until user responds]

1. Close: name the gaming-resistant design (multi-metric, audit, rotation, paired-constraint) + monitoring schedule.

[WAIT — do not advance until user responds]

The Process

Step 1 — State metric and goal: metric being targeted / underlying goal / current proxy-goal correlation / who is measured / stakes.

Step 2 — Predict the gaming: list ≥3 ways to game the metric with minimum effort on the goal. If you can't list 3, you haven't thought hard enough.

Step 3 — Categorize mechanism:

Step 4 — Choose countermeasure: Regressional → constrain range. Extremal → paired constraint metrics. Causal → closer-to-causation metric + direct audit. Adversarial → multi-metric + randomized audits + rotation. Step 5 — Design the system: primary metric / constraint metric(s) / audit mechanism (sampled direct goal observation) / rotation schedule / separation of measure-for-control from measure-for-diagnosis / gaming-detection threshold. Step 6 — Schedule re-evaluation: independent goal measurement (how/when/who) / drift threshold / retirement criteria / owner.

Output template

# Goodhart-Robust Design: 
Metric: | Underlying goal: | Correlation: | Who measured: | Stakes:
Gaming vectors (≥3):
Mechanism: Regressional / Extremal / Causal / Adversarial
Primary metric: | Constraint metric(s): | Audit: | Rotation: | Separation: | Gaming threshold:
Goal measurement (independent): | Drift threshold: | Retirement criteria: | Owner:

→ Method in Action: Goodhart 1975 (M3) and Strathern 1997 (RAE) → 2026 lens: AI benchmarks and engagement metrics as targets (2023–2026)

Pack: Goodhart's Law Patterns

Applying It Well

• All metrics are proxies — narrower than the goal. Pre-commit to gap analysis before deployment.
• Gaming is rational under measurement pressure. Fix the system, not the people.
• Rotation and audit are the only durable defenses. Plan metric retirement at design time.

→ Primary sources: references/sources.md

Common Rationalizations

[D] = designed upfront | [O] = observed in real use. [O] entries are more valuable.

Red Flags

• Metric tied to high-stakes rewards or punishments
• Metric "improves" without obvious improvement in the underlying goal
• People being measured can already articulate ways to game it
• Single metric is the primary evaluation tool, no audit or paired-constraint

Verification

• Goal underlying the metric specifically named
• Proxy gap explicitly described; ≥3 gaming vectors listed
• Goodhart mechanism category identified
• Countermeasure design (multi-metric, audit, rotation) in place
• Independent goal measurement scheduled with owner and retirement date

---

Part of deciqAI Knowledge Skills — 227 open-source thinking skills that make rigor executable for AI agents. The same skills power every deciqAI agent, which runs them autonomously to operate your company. See it run → https://www.deciqai.com/c/goodharts-law · ⭐ Star the repo → https://github.com/deciqAI/knowledge-skills · Contributions welcome.

Agents: latest version & machine-readable metadata → https://www.deciqai.com/s/goodharts-law.json