A KPI can improve while the outcome it stands for gets worse. Goodhart’s law names the risk: a statistical relationship tends to break down once you lean on it for control.
A KPI (key performance indicator) is a measure used to judge performance. Most KPIs are proxies: countable stand-ins for an outcome that is harder to count, such as customer trust or product quality. A measure reports how things are going. A target tells people which number to produce. Attach pay or pressure to the target, and people may find ways to move the proxy without moving the outcome.
So the question is not whether the number went up, but whether it still tells you what you need to know.
Who actually said “when a measure becomes a target”?
Goodhart made his point about monetary control in 1975. The sentence most people quote came from Marilyn Strathern in 1997.
Charles Goodhart, 1975. In his 2013 essay, Goodhart recalls that central banks in the early 1970s adopted monetary targets as an intermediate step toward controlling inflation. Each tended to choose the monetary aggregate that, until then, appeared to have the most stable relationship with nominal incomes, and hence inflation. By 1975 many of those relationships had broken down, and the breakdowns seemed most extreme for the aggregates chosen as targets.
That observation, he writes, led him to a comment at a Reserve Bank of Australia conference in Sydney in 1975: the breakdown accorded with “Goodhart’s Law, that any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes”. He adds that it “was intended as a humorous, throw-away line”. This follows the 2013 essay; the 1975 conference paper itself was not consulted.
Goodhart also points to something easy to miss: a new target “may alter not only the behaviour of the regulated but also that of the regulator”. The authority that set the target has credibility riding on meeting it, and may adjust its own behavior to get there.
Donald Campbell, 1976. The social scientist made a related argument about indicators such as test scores and police statistics: the more an indicator is used for social decisions, the more pressure there is to corrupt it, and the more apt it is to distort the activity it was meant to monitor.
Marilyn Strathern, 1997. The anthropologist’s article on audit in British universities contains the popular sentence: “When a measure becomes a target, it ceases to be a good measure.” Her example follows: “The more a 2.1 examination performance becomes an expectation, the poorer it becomes as a discriminator of individual performances.” A 2.1 is the upper second-class degree in the UK. She notes that Keith Hoskin describes this as “Goodhart’s law”. Goodhart, in turn, cites her wording in his 2013 essay.
| Who | Year | Field | The point |
|---|---|---|---|
| Charles Goodhart | 1975 | Monetary policy | A statistical regularity tends to collapse once it is used for control. |
| Donald Campbell | 1976 | Program evaluation | The more an indicator drives decisions, the more it invites corruption and distortion. |
| Marilyn Strathern | 1997 | Anthropology of audit | ”When a measure becomes a target, it ceases to be a good measure.” |
Goodhart’s version and Strathern’s make different points. Goodhart’s is about control: when you lean on a statistical relationship to steer by, it tends to give way, and the authority doing the steering may change as well. Strathern’s is about measurement: the more a result becomes the expectation, the less it tells performances apart. A KPI can fail in both ways at once.
None of this is a measured rate. Goodhart says his line rested on limited empirical observation, not deeper analysis, and nothing here tells you how often or how fast a KPI fails.
What does Goodhart’s law look like inside a company?
At Wells Fargo, a sales count drifted away from the customer need it was supposed to reflect.
In September 2016 the Consumer Financial Protection Bureau (CFPB), the US consumer-finance regulator, fined the bank $100 million. It wrote: “Spurred by sales targets and compensation incentives, employees boosted sales figures by covertly opening accounts”, funding them with money moved from customers’ own accounts without their knowledge or consent. According to the bank’s own analysis, employees opened roughly 1.5 million deposit accounts and applied for roughly 565,000 credit card accounts that, in both cases, may not have been authorized. The release adds that the bank would also pay $35 million to the Office of the Comptroller of the Currency and $50 million to the City and County of Los Angeles.
In February 2020 the US Justice Department announced that Wells Fargo would pay $3 billion to resolve criminal and civil investigations into a practice, running from 2002 to 2016, of pressuring employees to meet unrealistic sales goals. Its release describes millions of accounts or products provided to customers under false pretenses or without consent. In the agreed statement of facts summarized there, the bank admitted that its “onerous sales goals and accompanying management pressure” led thousands of employees to unlawful conduct, and that many of these practices were known inside the bank as “gaming”.
Three details in the Justice Department release matter for a KPI review:
- The proxy came loose from the outcome. The proxy was the count of products sold. The bank set out to sell existing customers more products while saying in public that it started “with what the customer needs”. Once pay and pressure were tied to the count, a higher count no longer showed a better-served customer.
- The check ran through the same hands. Employees altered customers’ contact information, the release says, both so customers would not learn of unauthorized accounts and so staff could not reach them for customer satisfaction surveys. A paired measure protects little if the people being measured can block access to its respondents.
- Leadership blamed individuals, not the model. An internal investigator called the problem a “growing plague” in 2004; another said in 2005 it was “spiraling out of control”. Senior leadership, the release says, cast it “as driven by individual misconduct instead of the sales model itself”. I read this as Goodhart’s point about the regulator, inside one bank.
The five-question KPI check
Pick one number used to pay, praise or rank people, and write down the outcome it is supposed to stand for. The middle column is my suggested standard; the Wells Fargo column uses only the cited releases.
| Question | What to write down | Wells Fargo, as documented |
|---|---|---|
| 1. What is it paired with? | A second measure, reviewed with the KPI, that could expose hidden harm. | Satisfaction surveys existed, but contact details were altered so customers could not be reached. |
| 2. Who can game it, and who checks it? | Who benefits from the number, who controls its records, and who can check them independently. | The employees opening unauthorized accounts could also change customer contact details. |
| 3. What behavior does it pay for? | The easiest action that moves the number without moving the outcome, in one sentence. | Opening products whether or not a customer needed or wanted them. |
| 4. What does it hide? | A cost or harm the number leaves out, and who bears it. | Fees the bank was not entitled to, harmed credit ratings, misused personal information. |
| 5. When do you retire it? | A review date, plus a trigger for revising or retiring it, such as “when it stops predicting renewals”. | Investigators warned in 2004 and 2005; leadership blamed individual misconduct, not the sales model. |
Treat a missing pair, a check the measured team controls, or a KPI with no retirement trigger as a gap to close before the next review. A second number does not help if the same people can move both. The questions help you test a KPI; they do not guarantee it stays useful.
Where does the cost of a gamed KPI land?
Not on the report that shows the KPI. The sales count recorded neither the customer harms nor the fine and settlement that followed in 2016 and 2020. That is what this site calls a downstream effect: a cost that arrives after the decision, on a surface the room was not watching, with no owner in the meeting.
DOWNSTREAM is my ten-step framework for tracing those costs before you commit. It comes from The Downstream Effects, which is in production and not for sale. It is an operating heuristic, not a validated risk instrument. The steps are Detect, Observe, Weigh, Navigate, Stream effects, Trace, Review, Effects on stakeholders, Assess systemic risk and Mitigate.
Step 8, Effects on stakeholders, names who actually pays. The book’s stakeholder map has five classes: customers, team, partners, regulators and future selves. Read against the cited releases, the sales goals give this map:
- Customers: fees and interest the bank was not entitled to collect, harm to some customers’ credit ratings, misused personal information.
- Team: thousands of employees drawn into unlawful conduct under sales pressure.
- Partners: the cited releases describe no separate harm to partners.
- Regulators and public authorities: the CFPB, the Comptroller of the Currency, the City and County of Los Angeles, the Justice Department, and the Securities and Exchange Commission (SEC).
- Future selves: the bank’s later settlement, including $500 million for the SEC to distribute to investors.
This map is drawn after the fact; it does not show that every consequence was predictable. Before you set a target, ask who could carry a cost the KPI leaves out. The full DOWNSTREAM framework covers the other nine steps.
Run the check this week
Block 15 minutes for one KPI. Write its intended outcome at the top of a page, answer the five questions, and mark what you cannot answer yet.
At your next team meeting, agree on the paired measure, name a checker whose pay or performance rating does not depend on the KPI, and put the review date and retirement trigger in writing. At the review, compare both measures and the records behind them with evidence of the outcome itself. If they disagree, have someone find out why before anyone calls the improvement a success.
For decisions that are not metrics, the second-order thinking audit runs four checks on later consequences.
Cite this:Goodhart's law: when a KPI stops measuring the outcome.Len P. van der Hof. https://lenvanderhof.com/en/blog/goodharts-law/ ·