Skip to content
Fritz Gerald ZephirinMenu

Social

Goodhart's Revenge

4 min read

The most useful sentence in management was written by an anthropologist studying British universities. In 1997, Marilyn Strathern compressed a decade of audit-culture damage into eleven words: when a measure becomes a target, it ceases to be a good measure. The pattern she named — now called Goodhart’s law — was already ancient. Soviet nail factories paid by weight produced absurdly heavy nails; paid by count, absurdly tiny ones. Wells Fargo targeted eight accounts per customer and got millions of fake accounts. Schools targeted test scores and got teaching-to-the-test. Wherever a proxy stands in for what you actually want, optimizing the proxy eventually divorces it from the thing.

For all of history, though, Goodhart’s law had a speed limit: humans game metrics slowly. Gaming takes effort, coordination, and a tolerance for shame, and most people game half-heartedly. The law was a chronic disease — metrics degraded over years, and you could often outrun the decay by updating them.

AI removed the speed limit. A large model is, mechanically, the most efficient target-hitting system ever constructed — that is what optimization is — and it hits targets without effort, without fatigue, and without the friction of knowing it’s gaming anything. Point it at a proxy and it will find the gap between the proxy and the intention faster than any human workforce, then exploit that gap a million times before your quarterly review. Goodhart’s law just went from chronic to acute. Three live examples, all from this site’s own archive:

Benchmarks. The industry’s measures of model intelligence became training targets, and benchmark contamination became a permanent scandal — every leaderboard inflates until it’s replaced, a succession of dead measures that is now just how the field works.

Support deflection. My home turf. Containment rate — conversations that never reach a human — became the industry’s headline KPI, and AI hits it magnificently: a bot that exhausts customers into giving up is contained. The metric reads victory; the churn data reads otherwise. Klarna’s public arc from “700 agents replaced” to “quality suffered, hiring humans again” is Goodhart’s revenge with a press cycle.

Engagement. Feeds optimized for attention got an infinite supply of content manufactured purely to capture it — an app made of nothing else hit #1 in two days. The metric is thriving. The thing it proxied is an open question.

And beneath all three, the human substrate Munger spent his life cataloging: incentive-caused bias means nobody near a gamed metric is motivated to notice. The AI optimizes the number; the team gets paid on the number; the dashboard summarizes the number upward. The machine and the org chart conspire, politely, to keep the proxy green.

So what survives? Not metric abolition — you cannot run an operation on vibes, and the alternative to measurement is whoever tells the best story. The discipline that works, which we run in production, is adversarial measurement design — assume every number you publish will be optimized against, because it will be:

Pair every proxy with its failure mode. Deflection ships with seven-day recontact and deflected-cohort satisfaction. Speed ships with error rate. Volume ships with a quality sample. A single metric is an instruction to game; a pair is a tension that has to be resolved honestly.

Keep an audit channel the optimizer can’t see. Raw transcripts, read by humans with power — sampling reality directly, off-dashboard. Elinor Ostrom found the same invariant in every commons that survived: monitoring by people with skin in the game, not just recorded rules. (Her book is the metric designer’s manual, disguised as political science.)

Rotate and retire. Treat metrics like credentials: the longer one stays a target, the less it measures. Mature numbers get demoted to context; fresh ones take the weight. Permanence is what Goodhart feeds on.

The strategic reading, since this is that kind of site: measurement just became a competency rather than a commodity. For seventy years you could run a company on borrowed KPIs, because everyone’s metrics decayed at the same slow human speed. Now the cost of hitting targets has collapsed while the cost of choosing targets — knowing what you actually want, and what would prove it — has not. That asymmetry should sound familiar by now. It’s the thesis of this whole archive, wearing its measurement costume: the optimizer is cheap. Knowing what to optimize is the job.

Sources & further reading