Goodhart's Law¶
Core Idea¶
Binding optimization pressure on a proxy degrades the proxy's correlation with the construct it was meant to indicate. Once consequences attach to a measure, agents take the cheapest path to move the proxy, which almost never coincides with the path that would move the underlying thing — and the proxy comes to index optimization effort rather than the construct.
How would you explain it like I'm…
Chasing the Sticker
Imagine your teacher gives a sticker for every page you read, hoping you'll learn more. Soon you start flipping pages super fast just to get stickers, without really reading. The sticker count goes up, but you're actually learning less — the thing the teacher really wanted got lost.
When the Stand-In Breaks
Sometimes you can't directly measure the thing you really care about, so you measure something easier that usually goes along with it — a stand-in. The trouble starts when you reward that stand-in, because people will then take the easiest path to push the stand-in number up, and that path is almost never the one that improves the real thing. So the stand-in number rises while the real thing you cared about quietly gets worse. It's not that measuring is bad or that everyone's cheating — even an honest person doing exactly what they're rewarded for makes this happen. Worst of all, the rising number is exactly what makes the problem hard to notice.
The Proxy Trap
Goodhart's law says that putting strong optimization pressure on a proxy breaks the proxy's connection to the real thing it was supposed to indicate. It needs three parts: a construct you actually care about that's hard or expensive to observe; a proxy that normally tracks that construct through some statistical regularity; and a control loop that attaches weight to the proxy — reward, penalty, promotion, funding. Once the loop closes, agents reallocate effort along the cheapest way to move the proxy, which is almost never the way that would move the construct, so the correlation that justified the proxy collapses and the proxy now measures 'effort aimed at the proxy.' This is sharper than 'measurement is bad' or 'people game systems' — it's the specific claim that using a regularity to control a system tends to destroy that regularity. And it doesn't need bad intent: a conscientious person doing exactly what the incentives reward widens the gap just as fast as a cheater, while the construct degrades behind the rising proxy.
Goodhart's law is the regularity that binding optimization pressure on a proxy degrades the proxy's correlation with the underlying construct it was meant to indicate. The pattern requires three elements working together: an unobservable or expensive-to-observe construct of interest; an observable proxy that, absent selection pressure, tracks the construct through some statistical regularity; and a control loop that places weight on the proxy — reward, penalty, allocation, status, promotion, regulatory consequence. Once the control loop closes, agents reallocate effort along the easiest path to move the proxy, which is almost never the path that would have moved the construct; the proxy-construct correlation that motivated the choice collapses, and the proxy now indexes 'optimization effort directed at the proxy' rather than the construct itself. The pattern is not 'measurement is bad' or 'people game systems' — it is the more specific, substrate-portable claim that the very act of using a statistical regularity to control a system tends to destroy the regularity. The mechanism is structural: any proxy capturing only part of a construct opens a wedge between proxy-improvement and construct-improvement, and selection pressure expands the wedge by preferentially recruiting the cheapest proxy-improvements, which sit disproportionately inside it. What distinguishes the law from generic measurement error, observer bias, or moral hazard is that the collapse arises from optimization directed at the proxy and does not depend on intent: Goodhart-collapse occurs from honest optimization as readily as from cynical gaming — a conscientious agent doing exactly what the incentive structure rewards widens the wedge just as surely as a cheater. The construct degrades while the proxy rises, and the rise is precisely what makes the degradation hard to see.
Broad Use¶
- Education: test scores drive funding and pay, so the easiest path becomes test-prep and narrowed curriculum while scores climb and learning does not.
- Healthcare: four-hour-wait and door-to-balloon targets distort triage and coding.
- Science: citation counts and p-values, tied to careers, breed citation cartels, p-hacking, and a replication crisis.
- AI alignment: trained policies maximize a reward signal in unintended ways — reward hacking is the canonical failure mode.
- Finance and monetary policy: targeting a monetary aggregate collapses the very regularity that motivated targeting it (the original case).
- Management: sales quotas and first-call-resolution metrics invite channel-stuffing and ticket-gaming.
Clarity¶
Distinguishes measurement (passive, low-stakes diagnosis) from incentivization (high-stakes, weighted into consequence): a metric can be excellent for one and catastrophic for the other.
Manages Complexity¶
Compresses Campbell's law, the McNamara fallacy, surrogation, reward hacking, and p-hacking into one mechanism, and bundles a portfolio of remedies — scorecards, held-out evaluation, rotation, decoupling — rather than a single domain fix.
Abstract Reasoning¶
Predicts which proxies collapse first (those with a wide, easily exploited wedge) and licenses robustness via multi-objective design — make the target a joint motion of several loosely correlated proxies so the wedge becomes a multidimensional intersection.
Knowledge Transfer¶
- Public administration: audit-and-investigate and multiple-measures accountability.
- AI: held-out reward evaluation and reward ensembling read as a re-derivation of those same remedies.
- Management accounting: the surrogation effect is the same proxy-collapse under a different name.
Example¶
An emergency department under a four-hour-wait target rushes disposition near the boundary, defers registration to delay the clock, and codes for compliance — the headline metric climbs while twelve-hour trolley waits and delayed-sepsis mortality fail to improve, exposing the proxy's collapse.
Relationships to Other Primes¶
Parents (1) — more general patterns this builds on
- Goodhart's Law is a kind of Proxy-Target Divergence — proxy_target_divergence's file states it directly: "Not goodharts_law. Goodhart is ONE decoupling mechanism — the agent-driven, strategic- adaptation one. This prime is the umbrella ... of which Goodhart is one child." goodharts_law's own file agrees it is the optimization-pressure-on- a-proxy mechanism. So goodharts_law is unambiguously a CHILD of proxy_target_divergence (valid candidate CAND-R25-006-06). High conviction; both files independently license it. Phase-C kept it OFF regulatory_capture (0.824 nearest, distinct mechanism), agency_problem, and moral_hazard — correctly. campbells_law is its high-stakes twin (both children of the same umbrella). NOTE: if the family is consolidated under candidate proxy_target_fidelity (the genus, see proxy_target_divergence EMERGENT), goodharts_law re-parents there; until then the umbrella is the built/ candidate target.
Path to root: Goodhart's Law → Proxy-Target Divergence → Proxy–Target Fidelity
Not to Be Confused With¶
- Goodhart's Law is not Regulatory Capture because Goodhart is a measurement-target failure needing no capturing party, whereas capture is a regulator co-opted by the industry it oversees; the fixes diverge (control-loop redesign versus structural independence).
- Goodhart's Law is not the Agency Problem because Goodhart bites even when objectives are perfectly aligned — the honest agent optimizing the proxy still degrades the construct — whereas the agency problem rests on a divergence of goals.
- Goodhart's Law is not Moral Hazard because Goodhart concerns a proxy's informativeness collapsing under optimization, independent of who bears risk, whereas moral hazard is changed risk-taking under insulation from consequences.