Goodhart's Law: When a Measure Becomes a Target, It Stops Being a Useful Measure

60-Second Summary

Goodhart's Law (1975): once a measure is used to evaluate or reward, it stops measuring what it used to.
Strathern's pithy reformulation (1997): 'When a measure becomes a target, it ceases to be a good measure.'
Real-world casualties: Wells Fargo accounts (2016), UK A&E 4-hour wait targets, Soviet nail factories, hiring quotas, engagement scores tied to bonuses.
Three failure modes: gaming, selection (only doing what's measured), and corruption (data falsification).
Fixes: rotate metrics, use unannounced sampling, pair lagging metrics with leading metrics, and never tie a single measure to compensation.

Your engagement score went up 12 points after you tied it to manager bonuses. Congratulations: you no longer have a useful engagement signal. The score now measures how well managers coach their teams to fill out the survey. That's Goodhart's Law, and it costs companies billions per year in misallocated trust.

Where it came from

Charles Goodhart was a Bank of England economist studying why monetary aggregates kept failing as policy targets. His 1975 paper argued: any statistical regularity will collapse once pressure is placed on it for control purposes. Anthropologist Marilyn Strathern's 1997 reformulation made it famous: 'When a measure becomes a target, it ceases to be a good measure.'

Three ways metrics rot

Goodhart failure modes

1
Gaming
People hit the number through technically-compliant behavior that doesn't deliver the underlying outcome. Soviet nail factory rewarded on tonnage → enormous useless nails. Rewarded on count → tiny useless tacks.
2
Selection
People only do what's measured. A&E rewarded on 4-hour wait times → patients held in ambulances outside (don't start the clock).
3
Corruption
When stakes are high enough, people fake the data. Wells Fargo opened 3.5M fake accounts to hit cross-sell targets. Fines: $3B+.

HR's worst Goodhart violations

Every one of these started as a real signal and became theater within a year of being tied to incentives.

Metric	How it rots	What it actually measures after 12 months
Engagement score tied to manager bonus	Manager coaches survey responses, suppresses critics	Manager's social pressure skill
Time-to-fill for recruiters	Lower bar, more passes through pipeline	Recruiter's willingness to push under-qualified candidates
Diversity hiring quotas	Reclassification + retention drop	Compliance theater, not actual inclusion
PR count for engineers	Tiny PRs, split commits	Skill at gaming the PR counter
NPS tied to CSAT bonus	Customer service agents beg for 10s, suppress detractors	Agent's begging frequency
Internal mobility rate target	Forced lateral moves with no intent to keep	HR's ability to fake movement

How to instrument without destroying the signal

Decouple measurement from reward where possible. Measure for diagnosis, reward on outcomes.
Use leading + lagging metric pairs. If engagement is the lagging metric, pair it with manager 1:1 frequency or skip-level attendance — both harder to game.
Rotate metrics quarterly so gaming patterns don't compound.
Use unannounced sampling instead of universal measurement — like the AET event-sampling approach.
Watch the second derivative. A metric that jumps 20% after being tied to a bonus is almost always being gamed, not actually improved.

The Wells Fargo lesson

Wells Fargo's cross-sell target (8 products per household) wasn't insane — it was based on the firm's then-best customers. The moment it became a target tied to firing, it produced 3.5M fake accounts, $3B in fines, and one of the largest reputation losses in US banking history.

Goodhart vs. Campbell

Donald Campbell's 1976 version (Campbell's Law — see the next article in this series) is the stronger statement: not just that metrics rot under pressure, but that the social processes they were meant to monitor get distorted too. Goodhart is about the measure; Campbell is about the underlying behavior the measure was meant to study. Both are true, and both apply to almost every HR dashboard.

FAQ

Frequently asked questions

Should we stop measuring engagement?

No — measure it, just don't tie it to bonuses. Diagnostic metrics survive Goodhart; reward metrics don't.

What about OKRs?

OKRs survive better than KPIs because they're explicitly aspirational and not directly compensated. They still rot when sneakily tied to comp.

Is there any metric Goodhart-proof?

Outcome metrics broad enough that gaming requires actually delivering the outcome — e.g. revenue per FTE, not 'calls per day'. But even those rot eventually.

Takeaways

Every metric you tie to money will become theater within a year.
Diagnostic vs. reward is the most important distinction in HR analytics.
If a metric jumped after you started rewarding it, you're probably measuring gaming, not improvement.