What is Goodhart's law?
Engineering culture and software practice
Goodhart's law holds that when a measure becomes a target, it stops being a good measure, because people optimise the number rather than the thing the number stood for. The economist Charles Goodhart made the point about monetary policy in 1975; the anthropologist Marilyn Strathern gave it its popular wording in 1997. In engineering it explains gamed velocity, coverage percentages, ticket counts, lines of code and, lately, AI usage metrics.
What this means
Goodhart's law describes what happens when a number that was useful for watching something quietly gets turned into a goal people are rewarded for hitting. As soon as that happens, people start working on the number itself, and the number drifts away from the real thing it was supposed to reflect. The measure was honest while nobody was trying to move it; the moment it carries consequences, it starts to lie.
The pattern has a name because it is so reliable and so easy to walk into with good intentions. A manager wants visibility, picks a sensible-looking metric, attaches a target to it, and is then surprised when behaviour bends around the target in ways that miss the point entirely.
The law does not say measurement is pointless. It says that measurement changes the thing being measured once it is used for control, so measures have to be designed with that in mind. The skill is keeping your eye on the purpose while using numbers as servants, not masters.
Why it matters
Any organisation that sets targets for its technical or AI work is exposed to Goodhart's law, and the exposure grows the harder the target bites. Tie a bonus, a league table or a management intervention to a single engineering number and you will almost certainly get the number, and quite possibly not the thing you actually wanted. The cost is not just wasted effort; it is decisions made on figures that have quietly stopped meaning anything.
It matters especially now, because AI programmes are awash with tempting numbers: prompts per employee, suggestion-acceptance rates, tokens consumed, seats activated. These are easy to count and easy to game, and counting them can crowd out the harder question of whether anything valuable is being produced. A dashboard that is green on all of them can sit on top of a programme that is delivering very little.
It matters for suppliers and service levels, where SLA targets invite gaming, and for internal reporting, where "dashboard theatre" lets everyone admire indicators that no longer track reality. Understanding the law lets a leader treat a suspiciously tidy metric with the scepticism it deserves, and design measurement that survives being used.
How it works
Where the term came from
Charles Goodhart, a British economist, made the original point in a 1975 paper written for a Reserve Bank of Australia conference, "Problems of Monetary Management: The UK Experience". His wording was technical: any observed statistical regularity, he wrote, "will tend to collapse once pressure is placed upon it for control purposes". He was explaining why the monetary measures the Bank of England relied on kept losing their predictive power once policy tried to steer by them.
The version everyone quotes came later. In a 1997 paper in the European Review on audit in British universities, the anthropologist Marilyn Strathern restated the idea as: "When a measure becomes a target, it ceases to be a good measure." That crisp sentence is what most people mean by Goodhart's law today. Two neighbouring ideas are worth a single mention: Campbell's law (1976), about the corruption of social indicators used for decision-making, and the Lucas critique in economics, both of which capture facets of the same problem. A 2018 arXiv paper by David Manheim and Scott Garrabrant, "Categorizing Variants of Goodhart's Law", sorts the effect into several distinct mechanisms.
Why measurement changes behaviour
A metric is a proxy: it stands in for something we care about but cannot measure directly, such as productivity, quality or value. The proxy holds together only while it is loosely coupled to the real thing. Attach a strong enough incentive and people find the cheapest way to move the proxy, which is almost never the same as improving the underlying reality. Effort flows to the number, the gap between number and reality widens, and the measure quietly stops being informative precisely because it is now being watched so closely.
The common engineering casualties
Software history is a graveyard of gamed metrics. Lines of code, used as a productivity measure, reward verbose, bloated software. Story-point velocity, once it becomes a target, gets inflated by generous estimation rather than more delivery. Test-coverage percentage becomes tests that touch code without checking anything. Ticket throughput encourages splitting and closing tickets rather than solving problems. In the AI era the same trap reappears with suggestion-acceptance rates, tokens consumed and prompts issued: all easy to inflate, none a reliable sign of value created.
The types of effect, lightly
Manheim and Garrabrant distinguish several routes to failure, including regressional effects (the proxy and the goal only correlated in the first place), extremal effects (pushing a measure to extremes breaks the relationship) and adversarial effects (someone deliberately games it). The practical lesson from the taxonomy is simple: there is more than one way for a target to corrupt a measure, so no single trick immunises you.
How experienced teams design measures that hold up
The durable response is to stop leaning on any single number. The SPACE framework, published by Nicole Forsgren and colleagues in ACM Queue in 2021, argues that developer productivity is multidimensional and cannot be captured by one metric, recommending measurement across several dimensions at once, using both hard data and human judgement. Practical habits follow from that: pair every quantity with a counter-metric that would get worse if the first were gamed; read numbers alongside qualitative review rather than instead of it; rotate which measures carry weight so none becomes a fixed goalpost; and keep asking what the metric is a proxy for, measuring the purpose rather than the proxy wherever you can.
Examples
A mid-sized firm launches an AI programme and, wanting to show adoption, sets a target of a certain number of prompts per employee per week. Within a fortnight the figure is hit handsomely, but a closer look shows staff firing off trivial or duplicated prompts to make the count, while the genuinely useful work the tool could do goes untouched. The measure that was meant to reveal adoption now hides the fact that little of value is happening.
A council department outsources part of its case-handling and writes a service level around the percentage of tickets closed within a target time. The supplier hits the target reliably, partly by closing tickets prematurely and reopening them under new numbers, and partly by reclassifying awkward cases out of scope. The residents whose problems the service exists to fix are no better off, but the dashboard is green.
A software team of eight is told that test coverage must reach a fixed percentage before release. Coverage climbs on schedule, but much of the new testing exercises code without actually checking that it behaves correctly, so defects continue to slip through. The number improved; the quality it was supposed to represent did not.
Common misunderstandings
A frequent misreading is that Goodhart's law means measurement is useless and should be abandoned. It does not. Goodhart himself was not against statistics; the warning is about turning a measure into a hard target, not about looking at data. Sensible measurement, held loosely, remains valuable.
Another is that gaming is always deliberate cheating by bad actors. Often it is not. Well-meaning people respond rationally to the incentive in front of them, and the measure degrades even when nobody intends to deceive. Treating it as a moral failing misses the structural cause.
A third confusion is with the myth of the 10x engineer. The 10x engineer idea is a claim about individual heroics and productivity myths, whereas Goodhart's law is not about who is talented; rather than ranking people, it is about how any measure corrupts behaviour once it becomes a target. Do not fold one into the other.
A fourth error is assuming that a better single metric will finally solve the problem. Because there are several distinct ways a target can corrupt a measure, no lone number is safe on its own; the fix is a balanced set, not a cleverer proxy.
A fifth is believing the law is only about money because it began in monetary economics. Its reach is general, which is exactly why Strathern's version, drawn from university audit, became the popular one.
Risks and boundaries
The law can itself be over-applied. Cited too freely, it becomes an excuse to avoid all targets and accountability, which is its own kind of failure; some measures under mild pressure behave perfectly well. The law bites hardest when the stakes attached to a single number are high and the number is easy to move directly.
There is honest debate about how far the effect generalises and how quickly it appears. Not every metric collapses the moment it is watched, and some proxies are robust enough to tolerate being targets for a long time. The taxonomy work helps here by making clear that "Goodhart" covers several different mechanisms rather than one, so blanket statements about it are usually too strong.
The boundary for a leader is judgement: the law is a reason to design and interpret measures carefully, not a reason to fly blind. Measuring nothing is not safer than measuring badly.
What to do next
Before you attach a target or an incentive to any engineering or AI metric, write down what the number is a proxy for and how someone could hit the number without delivering the real thing. If that shortcut is easy, choose a different measure or add a counterbalance.
Measure in balanced sets rather than single figures. Pair each quantity with a metric that would deteriorate if the first were gamed, so that improving both together is the only honest way to score well.
Always read numbers next to human judgement. A short qualitative review of what was actually delivered will catch dashboard theatre that no dashboard can.
Be especially wary of AI vanity metrics such as prompts, tokens, accepted suggestions or activated seats. Track value delivered and whether the work got genuinely easier, not activity that is trivial to inflate.
Rotate and retire measures deliberately. A metric that has become a fixed goalpost has probably already been gamed; changing what you emphasise keeps behaviour honest and keeps the numbers meaningful.
FAQs
Who came up with Goodhart's law?
The economist Charles Goodhart stated the original idea in 1975 in a paper on UK monetary policy. The popular wording, that a measure ceases to be a good measure once it becomes a target, is Marilyn Strathern's from 1997.
What did Goodhart actually say?
His original formulation, in 1975, was that any statistical regularity that has been observed tends to break down once it is used to steer policy. It was a point about monetary indicators before it was generalised.
Does the law mean I should stop measuring things?
No. It warns against turning a single measure into a hard target with strong incentives attached. Looking at data thoughtfully, and holding measures loosely, is still worthwhile.
What are the classic examples in software?
Lines of code, story-point velocity, test-coverage percentage and ticket throughput all degrade once they become targets, rewarding activity that moves the number rather than work that helps.
How does it apply to AI programmes?
Metrics like prompts per employee, suggestion-acceptance rates and tokens consumed are easy to game and easy to inflate, so a programme can look busy on the dashboard while delivering little of value.
How is this different from the 10x engineer idea?
The 10x engineer is a claim about individual productivity and heroics. Goodhart's law is about how any measure corrupts behaviour once it is targeted, regardless of who is involved.
What is the practical fix?
Use balanced sets of metrics with counter-measures, read numbers alongside qualitative review, measure the purpose rather than the proxy, and rotate measures so none becomes a permanent goalpost.