Skip to content
Mental Models

Goodhart's Law

Final Exam: Goodhart's Law

A graded, one-way final exam on Goodhart's Law — the claim, the decoupling mechanism and four flavours, the flagship cases, why it's a special case of incentives + principal–agent + feedback, its limits, and the fixes. Pass mark 70%.

20 min Updated Jul 10, 2026

This is the final exam for Goodhart’s Law. It pulls together the whole course: what the law claims and why proxies decay under pressure, the decoupling mechanism and its four flavours, the flagship cautionary tales, why Goodhart is really a collision of incentives, the principal–agent problem, and feedback loops, where the law does not apply, and the fixes that keep a metric honest. Take your time and reason each one through.

Warning:

How this exam works

Read carefully — this exam is final. Each question appears one at a time. Once you submit an answer it is locked for good: there’s no going back, no retry, and no restart. Your score is hidden until the end, where you’ll see a pass/fail verdict. The pass mark is 70%. A few questions ask you to select all correct answers.

Question 1 of 23

Which sentence best states Goodhart's Law, in Marilyn Strathern's sharpened form?

Select an answer to continue.

Course Recap

Big picture

Goodhart's Law, in one picture

  • Goodhart's Law
    • The claim
      • A measure made a target stops being a good measure
    • The mechanism
      • Borrowed correlation breaks; four flavours: regressional, extremal, causal, adversarial
    • Three parents
      • Incentives + principal–agent + feedback loops
    • Limits
      • Not anti-measurement; scales with stakes & gameability
    • The fixes
      • Baskets, loose targets, human audit, reward outcomes, rotate metrics
Success:

Key takeaways

When a measure becomes a target, it ceases to be a good measure. We reach for proxies because real goals are hard to see, and the proxy works only while it borrows a correlation with the goal — a correlation that optimisation pressure tears apart in four ways: regressional (you select the noise), extremal (you leave the range), causal (you push a lever that was never connected), and adversarial (a clever agent games you). Under the hood it’s incentives, the principal–agent problem, and feedback loops colliding. But Goodhart is not a ban on measuring: it’s a warning to hold targets loosely, use baskets of hard-to-game metrics, keep humans auditing, reward real outcomes over proxies, and rotate metrics as they decay. Do that, and you can keep measuring the world without letting the world game you back.

Mark lesson as complete