Goodhart's Law: The Tyranny of Metrics

Imagine you are the CEO of a software company. You are an idealist and a modern individual, strongly against the pre-modern management approach we discussed last time, and you are willing to trust strangers. You are determined to build a fair, reasonable, and meritocratic system. But with hundreds of employees, how can you know who to promote and reward?

To ensure objectivity and fairness, you set a numerical performance metric for the engineering team: who fixes the most bugs.

Initially, the results were indeed excellent; you watched the number of bug fixes shoot up rapidly, and you were very satisfied.

However, after only a few months, something felt off. HR actually suggested “optimizing out” Old Zhang from the team. But you know Old Zhang—he’s the best engineer, and he built the company’s most difficult core architecture years ago. HR’s reason was that Old Zhang fixed the fewest bugs. You specifically inquired and realized that the area Old Zhang was responsible for simply barely had any bugs; how could he fix them?

And Little Li was promoted to supervisor because he fixed the most bugs. But upon closer inspection, you found something wrong: Little Li specifically picked easy, minor bugs, closing several in a day; he often broke down one large task into five smaller work orders and reported them separately; but when faced with tough nuts to crack, he would tag them as “pending” and leave them there. Little Li’s fix count was equivalent to half the team, but was he really that useful?

If this metric-based assessment continues, the entire company will learn from Little Li, and no one will want to be Old Zhang.

This is Goodhart’s Law.

Goodhart’s Law is a special case of “unintended consequences,” meaning that a metric distorts its original intention—a metric is originally used to observe reality; but once it becomes a target for reward or punishment, people will start optimizing the metric, rather than optimizing reality itself.

The Origin and Core Mechanism of Goodhart’s Law #

Charles Goodhart is a British economist. In 1975, while discussing UK monetary policy, he made a rather convoluted statement to the central bank, the gist of which was: Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes [1].

This idea was later refined by anthropologist Marilyn Strathern in 1997 into a popular adage—

“When a measure becomes a target, it ceases to be a good measure.” [2]

The process described by Goodhart’s Law is essentially as follows—

First, you have a “true objective,” which is what you genuinely want, possibly health, knowledge, service quality, or the company’s real performance. But this true objective is often complex and difficult to quantify, so you propose a “proxy indicator,” which is a number you use to approximate that true objective. These are weight, test scores, paper counts, bug fixes, and likes. You cannot directly assess the true objective, so you specifically assess the proxy indicator, and the assessed individuals then face “optimization pressure,” beginning to strive to make that indicator look good… Ultimately, no one cares about your true objective anymore.

Goodhart’s Law has two close relatives. One is the Lucas Critique, which we discussed earlier, specifically for economic policy, meaning that people automatically optimize their behavior in response to new policies, so you cannot use past relationships to predict the effects of new policies [3]. The other is called Campbell’s Law, for social phenomena, meaning that the more a social indicator is used for major decisions, the more susceptible it is to corruption and the more it will distort what it was originally intended to measure [4].

We can also compare Goodhart’s Law with legibility discussed earlier: legibility is the state distorting society for administrative convenience; Goodhart’s Law is society actively distorting itself to cater to management.

Manifestations of Goodhart’s Law in Reality #

Once you understand this pattern, you will find Goodhart’s Law everywhere.

Perhaps the true objective of internet platforms is to provide you with good service, but they treat “time spent” and “likes/shares” as proxy indicators of your “liking.” The result is that what keeps you hooked is often not what’s best for you, but what’s most addictive and emotionally provocative. Research suggests that the more recommendations are sorted by engagement signals, the more they amplify misinformation and extreme polarization [5].

The UK’s National Health Service (NHS) wanted to shorten emergency wait times and set a strict target for hospitals: “must be treated within 4 hours.” Hospitals learned to queue patients in ambulances and leave them in corridors—the timer wouldn’t start until the person “officially arrived”; when it approached 4 hours, they would quickly admit the patient to stop the clock [6].

Similarly, if police are assessed by their case clear-up rate, some will specifically target easy, minor cases, leaving time-consuming major cases aside. If government departments are assessed by “budget execution rate,” they will rush to spend money by year-end…

Goodhart’s Law is the inevitable fate of numerical governance in complex systems.

This is even true for training AI.

We want AI to act according to human preferences, so one way to train large models is to first create a “robot judge” that imitates human likes and dislikes to automatically score the model’s performance. During training, the model will instinctively strive to please this judge. But this judge is only an imperfect proxy for human preferences! Research has found that if large models significantly over-optimize the scores received from the “robot judge,” their true, human-useful capabilities actually decline—the researchers directly wrote in their paper: this is “exactly consistent with Goodhart’s Law” [7].

As long as there’s an assessment, even AI can learn to game the system.

Academia: A Hotbed of Metric Worship #

I have to say, the area most severely distorted by Goodhart’s Law is academia. Scholars worldwide value publishing papers, but China, in particular, treats papers as almost the sole assessment metric. Papers are no longer academic exchanges for peers; they are medals for promotion and funding, and even become piecework.

A study surveyed 168 documents from 100 Chinese universities and found that during the era of peak “SCI worship” from 1999 to 2016, these universities offered direct cash rewards for papers indexed by Web of Science, with bonuses ranging from the equivalent of $30 to $165,000, with the highest being approximately 20 times a professor’s annual salary [8].

If publishing papers is piecework, you can imagine that there will be a large number of low-quality papers, and even fake papers. China’s paper output has long been the world’s number one [9], but at the same time, the number of retractions is also the world’s highest, accounting for more than half of the global total [10].

In 2020, the Ministry of Education and the Ministry of Science and Technology, perhaps unable to bear it any longer, jointly issued a document, admitting the “partial, excessive, and distorted use” of SCI paper-related indicators in current research evaluation, leading to “an alienated phenomenon of making the number of SCI papers, high impact factor papers, and highly cited papers the ultimate goal,” and demanded the cancellation of direct rewards based on SCI indicators for individuals and departments [11].

Such documents will not have much effect, because metrics can evolve.

Initially, everyone looked at the number of SCI papers. Later, people realized that the quantity of papers did not equate to quality, so they began to focus specifically on citation counts. Later, more advanced citation metrics emerged, such as the h-index. When it was found that a pile of low-quality papers citing each other was also meaningless, various institutions began to chase “harder top-journal metrics”: the number of publications in CNS (Cell, Nature, Science, the three universally recognized top journals), or top-tier journal combination rankings like the Nature Index.

Now, Chinese researchers have also gamed their way to world number one in the Nature Index rankings [12]… yet China, since its reform and opening-up, has still not produced a scientific discovery worthy of a Nobel Prize.

That’s right, as long as you set a metric, people will game it. And once people start gaming it, it ceases to be a good metric.

Deep Reasons for Metric Worship: Insufficient Authority and Lack of Trust #

So why must we resort to metric governance? Can’t people see Old Zhang’s high caliber without metrics?

In fact, skill is not some mysterious thing. Take scientific research: can you open up a new direction? Have you solved real problems? What is the quality of your work, even whether you seriously mentor students, whether your questions are insightful—these are clear in the eyes of insiders. As the saying goes, talent is like pregnancy, it cannot be hidden.

Never underestimate human judgment. People know good from bad. Why can’t Nobel Prizes be gamed? Because the Nobel Prize is based on human judgment, not metrics; there are no standard answers, it’s not about crossing a certain threshold to receive an award.

Human judgment is sophisticated precisely because it relies on hard-to-quantify “tacit knowledge” with no fixed standards. If you produce such groundbreaking research today, it is earth-shattering; if you publish a similar paper next year, it will be worthless.

The traditional method in academia is to respect human judgment, practice peer review, and have faculty governance, where insiders decide who gets promoted to full professor and who receives funding. But in the eyes of power, this is too uncontrollable.

Administrative power seeks control, but you can’t just arbitrarily promote anyone to professor—how would you gain acceptance? Using metrics is the best approach.

American science historian Theodore Porter’s 1995 book, Trust in Numbers [13], specifically addressed this issue. Porter found that those most obsessed with quantification are often precisely the institutions lacking sufficient authority.

A fully authoritative expert can directly decide, “I believe he is capable,” and no one dares to question. But a bureaucrat lacking confidence, fearing dissent, or shirking responsibility, would not dare to do so. Thus, they resort to numbers—numbers appear objective, impersonal, so no one can attribute blame to them. Porter states that quantification is a way to “make decisions while pretending that no one is making decisions”; and objectivity, precisely, lends an aura of authority out of thin air to officials who inherently lack it.

So the root cause of Goodhart’s Law is not too strong power, but too weak power; not too strict control, but incompetence and lack of authority to control.

Metric worship is not a victory for rationality, but a failure of judgment.

The weaker the boss, the more detailed the KPI.

Metrics and Judgment: How to Balance? #

But then again, relying entirely on human judgment isn’t necessarily reliable either. Academia can easily devolve into cliquishness, nepotism, and mutual back-patting. The original meaning of the term “academic bully” (学霸) refers to those who monopolize academic resources, control discourse, and suppress dissent.

Metric tyranny and black-box human governance are merely two ends of a spectrum: judgment must have standards, otherwise it is arbitrary; but once standards are fixed, they become targets for optimization and circumvention.

Behind Goodhart’s Law lies the inherent tension between professional judgment and procedural justice… there is no perfect selection and review system in the world.

But this does not mean that slightly better selection and review systems do not exist.

Strategies and Principles for Countering Goodhart’s Law #

In fact, Goodhart’s Law has never opposed quantification; it opposes naive quantification. Numerical metrics can certainly serve as one basis for judgment, but they should not be the judgment itself.

A good evaluation system should be like a courtroom: it certainly emphasizes procedural justice, but it will not stipulate “three witnesses mean conviction” or “the party with more evidence wins”—it values evidence and procedure, but allows for a degree of discretionary power.

Take academic evaluation, for example; there are already some exploratory reforms underway, with quite consistent directions.

In 2015, several international bibliometricians published The Leiden Manifesto in Nature, whose first principle clearly states: quantitative evaluation can only support, not replace, expert qualitative judgment [14]. The UK’s Research Excellence Framework (REF) is a more institutionalized example: it is a periodically conducted national university research assessment in the UK, and its results influence the allocation of university research funding—it explicitly forbids using journal impact factors to substitute for judgments of paper quality, instead focusing on a limited number of representative works, genuine social impact cases, and descriptions of the research environment [15].

In addition to prevention, you should also engage in post-hoc reflection to see which metrics have been “Goodharted.” One study found that if managers genuinely participate in “strategic choice”—meaning personally participating in judging what the company truly wants to pursue, rather than just receiving a metric sheet to execute—they are less likely to mistake metrics for ends [16].

Based on these experiences, the methods to counter “Goodhartization” primarily follow three principles—

First, metrics serve as input, not verdicts. Citation counts, journals, awards can all be laid on the table for reference, but there should be no rule like “X papers equals promotion, otherwise a veto.” In short: metrics are witnesses, not judges.

Second, evaluate by representative works, and this must be accompanied by written justifications. Don’t count the total number of papers an individual has published, but look at their best three to five pieces of work—and force reviewers to clearly write down: what is this person’s core contribution, what problem did they solve, and what are the dissenting opinions? Counting papers is about quantity; reading representative works is about quality.

Third, categorize evaluation by role. University teachers, clinicians, and engineers are inherently three different jobs and should not be flattened by a single standard.

On top of these three principles, add several procedural safeguards: review criteria must be public, stakeholders must recuse themselves, and unsuccessful candidates must have a place to appeal… and periodically conduct “anti-audits” every few years—to check if this set of metrics has induced any bad behavior, and if so, change them.

People require institutional management, but institutions are dead, people are alive, and people can modify institutions. Management is not installing a system; it is tending to a system.

Everyone Can Become a “Digital Slave” #

We must accept that Goodhart’s Law will always exist.

As long as you desire objectivity and fairness, rewards and punishments must have standards → others will adapt to these standards → these standards will be gamed → requiring you to revise the standards. The Goodhart and anti-Goodhart process will never end.

The scariest part is that even without a boss to manage you, or metrics to constrain you, you will Goodhart yourself.

Haven’t you, perhaps unconsciously, equated “hours spent studying” with learning, “the number on the scale” with health, “leaving late” with merit, “thousands of words written today” with creation, “instant replies” with love?

Without realizing it, dating activities have devolved into a competition of age, income, height, education, and dowry; travel has become checking off sights, photos, and step counts; even filial piety is quantified by how much money was transferred and how many times one visited home. People live their lives as a pursuit of numbers, turning themselves into living spreadsheets—no one forces you, you willingly relinquished your power of judgment.

At that point, metrics have become your master.

As the poem attests:

A ruler tries to gauge the sky, but sky is not a measure. A map may guide, but roads aren’t lines, nor map a true treasure. Though panels glow green, the machine still decays inside. Though scales report a lighter frame, the body’s life has dried. Weak power clings to myriad rules, timid minds seek simple signs. Know metrics are but servants true, let not your life become their lines.

Notes

[1] Goodhart, Charles A. E. “Problems of Monetary Management: The U.K. Experience.” 1975.

[2] Strathern, Marilyn. “‘Improving Ratings’: Audit in the British University System.” European Review 5, no. 3 (1997): 305–321.

[3] Lucas, Robert E., Jr. “Econometric Policy Evaluation: A Critique.” Carnegie-Rochester Conference Series on Public Policy 1 (1976): 19–46.

[4] Campbell, Donald T. “Assessing the Impact of Planned Social Change.” Evaluation and Program Planning 2, no. 1 (1979): 67–90.

[5] Germano, Fabrizio, Vicenç Gómez, and Francesco Sobbrio. “Ranking for Engagement: How Social Media Algorithms Fuel Misinformation and Polarization.” Barcelona School of Economics Working Paper, 2025.

[6] Bevan, Gwyn, and Christopher Hood. “What’s Measured Is What Matters: Targets and Gaming in the English Public Health Care System.” Public Administration 84, no. 3 (2006): 517–538.

[7] Gao, Leo, John Schulman, and Jacob Hilton. “Scaling Laws for Reward Model Overoptimization.” arXiv:2210.10760, 2022.

[8] Quan, Wei, Bikun Chen, and Fei Shu. “Publish or Impoverish: An Investigation of the Monetary Reward System of Science in China (1999–2016).” Aslib Journal of Information Management, 2017.

[9] Institute of Scientific and Technical Information of China. “2024 Statistical Report on Chinese Scientific and Technological Papers,” 2024.

[10] Xu, Shuang, and Guangwei Hu. “Reckoning with Retractions in Research Funding Review: The Case of China.” Publications 13, no. 3 (2025): 41. See also Van Noorden, Richard. “More than 10,000 Research Papers Were Retracted in 2023 — a New Record.” Nature, December 12, 2023.

[11] Ministry of Education, Ministry of Science and Technology. “Several Opinions on Regulating the Use of SCI Paper-Related Indicators in Higher Education Institutions and Establishing Correct Evaluation Guidance” (Jiao Ke Ji [2020] No. 2), February 2020.

[12] Nature Index. “2025 Research Leaders: Leading Countries/Territories.” Springer Nature, 2025. https://www.nature.com/nature-index/research-leaders/2025/country/all/global/all.

[13] Porter, Theodore M. Trust in Numbers: The Pursuit of Objectivity in Science and Public Life. Princeton University Press, 1995.

[14] Hicks, Diana, Paul Wouters, Ludo Waltman, Sarah de Rijcke, and Ismael Rafols. “Bibliometrics: The Leiden Manifesto for Research Metrics.” Nature 520 (2015): 429–431.

[15] Research Excellence Framework (REF 2021). “Panel Criteria and Working Methods.” 2019.

[16] Choi, Jongwoon (Willie), Gary W. Hecht, and William B. Tayler. “Strategy Selection, Surrogation, and Strategic Performance Measurement Systems.” Journal of Accounting Research 51, no. 1 (2013): 105–133.