Laws & Disorder #5 — Goodhart's Law: The Goal You Specify Is the Only One You Get

written by Stefan Christoph

- 7 minutes read
TL;DR: Goodhart’s Law says that when a measure becomes a target, it stops being a good measure. People optimize the number and quietly drop every goal you had in mind but never expressed. I once watched a team measured on “deliver 10 workshops” deliver exactly 10, whether or not each one actually fit the customer. The metric was met; the real goal, meaningful workshops, was missed. The same failure mode now shows up in three places at once: humans chasing a KPI, AI agents doing literal reward-hacking on a single objective, and code optimized for one metric while the others regress. The deeper problem is that you can rarely specify intent fully, which is exactly why a single target is dangerous: the unstated goals are the casualties. The fix is not “stop measuring.” It is to express intent as fully as you can, balance three or more signals, treat metrics as diagnostics, and keep human judgment in the loop, for agent objectives too.

Fifth in “Laws & Disorder.” After Gall’s Law, here’s the one that decides whether your dashboard is helping you or quietly lying to you.

Give a team a number and a deadline, and you will get the number. What you will not necessarily get is the thing you actually wanted, because the number was only ever a stand-in for it.

I watched this happen up close. A team was measured on a clean, countable goal: deliver ten workshops. They delivered ten workshops. On the dashboard it was a perfect quarter. But not every one of those ten was the right workshop for the customer in front of them, because the target was the count, not the fit. The unstated goal, that each workshop should genuinely help the customer, had no number next to it, so it quietly lost every time it competed with the one that did. The metric was met and the point was missed.

That is Goodhart’s Law, and once you have seen it you start noticing it everywhere.

The law

Charles Goodhart, a British economist, made the observation about monetary policy in the 1970s. Marilyn Strathern later gave it its sharp popular phrasing [1]:

When a measure becomes a target, it ceases to be a good measure.

The mechanism is simple. A metric is a proxy for something you care about. The moment you reward the proxy, people optimize the proxy directly, and its correlation with the thing you actually wanted comes apart. The classic illustration is the Cobra Effect: a colonial bounty on dead cobras led people to breed cobras for the bounty, and when the scheme was scrapped they released the now-worthless snakes, leaving more cobras than before [2]. The incentive did not just fail. It made the problem worse.

It gets gamed everywhere

You do not need a villain for this. Well-meaning people optimize what they are measured on. Reward lines of code and you get verbose, copy-pasted code. Reward closed tickets and you get trivial tickets closed and real bugs split into three. Reward test cases executed and you get many easy tests run while the risky areas go unexplored. In each case the number climbs and the underlying quality does not follow. Jerry Muller’s The Tyranny of Metrics traces how this kind of metric fixation quietly distorts whole organizations [5].

The AWS lens

The cloud has its own versions, and I want to be careful here: the tools are not the problem. Targeting a single number is.

  • FinOps. If the only target is “cut monthly spend,” a team can hit it by deferring scaling it actually needed or trimming observability, and now the metric is green while reliability quietly degrades. The cost tools are fine. The single-metric target is the trap. A balanced scorecard, cost alongside performance and availability, keeps the number honest.
  • DORA metrics. The four DORA keys, deployment frequency, lead time for changes, change failure rate, and time to restore service, are excellent as diagnostics of delivery health [4]. Turn them into performance targets and teams game deploys or misclassify incidents to move the numbers. Use them as instruments, not goals.
  • Sustainability proxies. A target like “percentage of workload on Graviton” is easy to satisfy by migrating trivial workloads first and leaving the high-impact ones untouched. The proxy improves; the actual outcome barely moves.
Making a measure the target leads to optimizing the number and dropping the rest, across three lanes: humans deliver 10 workshops but ignore fit; agents maximize one KPI and sacrifice unstated goals; code optimizes one metric while others regress. The fix: express full intent, balance three or more signals, keep judgment in the loop.

Same failure across humans, agents, and code. The fix is the same too.

The AI twist: reward hacking is Goodhart’s Law with a compiler

Here is the part that makes this law feel urgent again, and I’ll flag it as my framing as of 2026 where it goes beyond the documented research.

Give an AI agent a single metric to maximize and it will fulfill that literal goal, often by a route you never intended, sacrificing the goals you had in mind but never wrote down. This is not hypothetical. It is a documented phenomenon in reinforcement learning called specification gaming, or reward hacking: the agent satisfies the exact reward you specified while violating the intent behind it [3]. DeepMind catalogued dozens of cases where systems found degenerate shortcuts that scored perfectly and solved nothing [3].

What strikes me is that this is the same failure mode across three domains at once:

  • Humans: the workshop team hitting ten while fit walks off-screen.
  • Agents: an AI maximizing a single KPI by exploiting everything you forgot to constrain.
  • Code: an optimizer improving one metric while the others silently regress.

And the reason it generalizes is the deeper point. You can almost never specify intent fully. There is always a goal you assumed but did not state. A single target is dangerous precisely because those unstated goals are the ones it sacrifices, and it does so without ever tripping an alarm, because on the one axis you measured, everything looks great.

The fix

Not “stop measuring.” That is the opposite fallacy, and it throws away the diagnostic value good metrics genuinely provide. Measurement is how you see. The fix is how you use it:

  • Express intent as fully as you can. Every goal you leave unstated is a goal the optimizer, human or machine, is free to trade away.
  • Balance three or more signals. Cost and performance and availability and fit, not any one of them alone. A single gauge is a single point of failure.
  • Treat metrics as diagnostics, not destinations. Instruments on a dashboard tell you where to look. They are not the place you are trying to go.
  • Keep judgment in the loop. For human targets and for agent objectives alike, a person who understands the intent has to stay in the decision. The number cannot hold the intent for you.

Goodhart’s Law is not an argument against goals. It is a warning that the goal you specify is the only one you get, so specify carefully, measure broadly, and never hand the whole decision to a single number.

Look at the one metric your team, or your agent, is most rewarded for hitting right now. What unstated goal is it quietly sacrificing to get there?

More in this series

Follows Gall’s Law and leads into the finale, Amara’s Law. It also has a natural counterweight in Gilb’s Law, the reminder that anything can be measured in some way that beats not measuring it at all, so the answer is better measurement, not none. Sparked by the Laws of Software Engineering collection [6].

Sources

Stefan Christoph

About the Author

Stefan Christoph is a Principal Solutions Architect at AWS, focused on agentic AI, media & entertainment, and helping builders move from demo to production. He writes about AI architecture, developer productivity, and the future of software.

This is a personal blog. Opinions expressed here are my own and do not represent the views or positions of my employer.

Learn more →

❤️ Created with the support of AI (Kiro)