The Wager Inside Every Measure

In 1975 a monetary adviser at the Bank of England named Charles Goodhart wrote down what he kept watching happen to his own instruments. Every statistical signal the Bank found useful stayed useful right up until the Bank steered by it — and then it went quiet. His sentence was dry as paper: any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes. Twenty-two years later the anthropologist Marilyn Strathern compressed it into the form everyone now carries: when a measure becomes a target, it ceases to be a good measure.

He was not alone. Within about a year of his paper, Donald Campbell said the same thing about social indicators — the more a quantitative indicator is used to decide things, the more it corrupts and is corrupted — and Robert Lucas said it about entire economic models. Three fields tripped over the same stone in the same stretch of road. When that happens, the stone is not a quirk of one discipline. It is structural.

01The clause Goodhart left out

But the law, as usually told, is only half a law. It explains why targeted measures die. It does not explain why so many of them lived so long. The diploma was a target for a century and mostly held. The cover letter was a target for decades and mostly held. If targeting alone killed measures, they should all have died the day they were posted on the wall.

They held because every working measure is a wager: a bet that the cheapest way to move the measure is to do the thing it measures. The diploma held because sitting in the rooms for four years cost nearly as much as getting the education. The cover letter held because writing clearly cost nearly as much as thinking clearly. The confident answer in the interview held because fluency about work you had never done was genuinely hard to fake for an hour.

The measure was never the thing. It was collateral­ized by the price of faking the thing — and nobody signed that wager, so nobody noticed it could be repriced.

02The repricing

AI moves exactly one side of that bet. It does not change what it costs to be able to do a job. It changes what it costs to produce the symbols of being able to — and for any measure made of symbols, words, scores, portfolios, benchmarks, applause, the new price of the fake is approximately zero.

Goodhart’s law used to run at the speed of human effort. Teaching to the test took a semester. Padding a résumé took an evening each time, per applicant, and shame was a surcharge. Because gaming took work, measures died one at a time, over years, and every institution could tell itself the problem was fraud — a few bad actors — rather than arithmetic. That story is now over. Every symbolic measure shares the same collateral, and the collateral was just marked to zero. The collapse is not sequential anymore. It is simultaneous.

Notice what actually breaks. A measure is an artifact deputized to stand for an Actual — for something that occurred. Game the measure and you manufacture the artifact without the state change it was deputized to report. The forged artifact still enters the record, and the record is what prediction trains on: yours, your employer’s, every machine’s. Nobody is touching any denominator; that road stays closed. What is being poisoned is the slow, lawful road itself. Expectation keeps learning from a record that is filling with events that did not occur.

03Where measurement goes

Measurement does not die when the symbols die. It retreats to the one ground that cannot be manufactured: what has already happened. Here is what I think that looks like.

1. The symbolic filter is abandoned, not reformed.

   Résumé screens, cover letters, admissions essays, unproctored take-homes. Each gets one round of detection theater, and then quietly stops being read. You cannot patch a measure whose collateral is gone. You can only stop lending against it.

2. Dashboards will look their best just before they are turned off.

   Expect a short golden age of gamed metrics: every number up, nothing underneath moving. The institutions that keep steering by symbols will be outcompeted by the ones that steer by settlements, and the difference will show up before the dashboards admit anything.

3. Measurement moves to after.

   A score is a prediction; a settlement is a record. Outcome pricing, guarantees, pay that follows the result — the measure of work becomes the past-tense verb, checked once it is checkable. Before-the-fact measurement shrinks to the cases where you truly cannot wait.

4. The room comes back as an instrument.

   When every symbol can be manufactured, the shrinking set of expensive things gets repriced upward — and near the top of that set is a person, present, doing it now. Interviews get shorter, rarer, and physical. Presence becomes part of the measurement stack again, not a courtesy.

5. Stakes replace scores.

   Where an outcome cannot be waited for, the substitute is not a better test. It is someone underwriting you: deposits, bonds, warranties, an insurer pricing a signature. Where symbols fail and time is short, skin arrives.

04The claim ledger

Already true

Machines produce interview-grade answers, essay-grade prose, and portfolio-grade artifacts that are indistinguishable at reading distance. Application volume has outrun human reading in most large hiring funnels, so machines are already screening symbols that machines wrote. The wager is already repriced; the institutions are what lag.

What has to happen

Outcome verification has to get cheap and legible enough to lend against — paid work samples, sandboxed trials, guarantee pricing, pay-after rails — before the last symbolic filters die. If it does not arrive in time, hiring and admissions fall back to the oldest proxy of all: who you already know. The window matters more than the destination.

Where I am probably wrong

The sealed room. Identity verification plus proctoring can rebuild the collateral by force: lock a verified human in an enclosure and the symbol becomes expensive again. If verifying a person gets cheap faster than verifying an outcome, the future of measurement is not the settled record — it is the exam room, everywhere. That would falsify half of this piece, and it is not a fringe possibility.

The résumé is still open on your laptop. It is past eleven, and the filter you are performing for is one of the last of its kind — this year it may still be read, so finish the translation if you need to. But the wager underneath it has already been repriced, and every institution will eventually notice what you noticed tonight: the symbol and the substance have finally, completely, come apart.

So spend the better hours on the other side of the split. Do the work that leaves a record no one has to take your word for.

Go make it past tense.