This is the failure that has a name, two names in fact, and it is the reason the other six articles in this section exist.
Charles Goodhart, writing about monetary policy in the mid-nineteen seventies, observed that a statistical regularity tends to collapse once pressure is put on it for control purposes. Donald Campbell, writing at much the same time about social indicators, put it more bluntly: the more an indicator is used for decision-making, the more it will distort and corrupt the processes it was meant to monitor. Both are academic sources with nothing to sell, which is why they are worth naming rather than paraphrasing as folk wisdom.
Why the collapse happens
A proxy correlates with the thing it stands for because, historically, the easiest way to move the proxy was to move the thing. That correlation was measured in a world where nobody was trying.
Attach a consequence and people start trying. Trying means searching the space of everything that moves the number, and that space is much larger than the space of things that also move the target. Most of the routes it contains are cheaper than doing the work, which is precisely why they get found.
The correlation does not weaken because people are dishonest. It weakens because it was never a causal law, and optimisation is very good at finding that out.
What it looks like in each market
On the systems side: an availability target met by narrowing what counts as an incident. A latency target met by returning an error quickly, since a fast failure is faster than a slow success. A ticket-resolution metric met by closing and reopening. A deployment frequency target met by splitting one change into six.
On the workforce side the routes are equally well known, and one of them has a supply chain: devices sold openly for the sole purpose of moving a mouse pointer so that activity monitoring records presence. Their existence as a commercial product category is the clearest available evidence of the mechanism, because a market only appears where there is demand, and the demand is created by the measurement.
None of these require anyone to decide to cheat. They are what the incentive selects for, and a system that punishes the person who does not use them is selecting for them deliberately.
Better metrics do not solve it
The instinct after being gamed is to find a cleverer measure. It buys time. Any single number is optimisable, and the effort spent finding the route scales with the size of the consequence attached.
What helps is structural rather than clever.
Pair every metric with a counter-metric. Speed with quality. Throughput with rework. Utilisation with error rate. Activity with delivered outcome. A pair is much harder to move together by any route that does not involve doing the work, and where somebody manages it anyway, they have probably found a genuine improvement.
Measure as close to the outcome as the system allows. The distance from the proxy to the thing, from the first article in this section, is exactly the size of the gaming surface. Requests served correctly to a real user is a smaller surface than processes alive.
Keep a human judgement in the loop for consequences. Numbers are evidence in a decision, not the decision. An organisation that automates the step from measurement to consequence has removed the only component capable of noticing that the measurement stopped meaning anything.
Review the measure on a schedule. Nothing announces that it has been gamed. The signal is usually a number improving while the thing it stood for does not, which requires somebody to be looking at both.
The transparency tension, stated honestly
Telling people exactly what is measured is the right thing to do, is required in many jurisdictions for workforce monitoring, and improves adoption. It also tells them precisely what to optimise.
Both halves are true and the trade is real. Concealing the metric does not prevent gaming, it delays it and produces resentment in the interim, and covert measurement carries the legal exposure covered in the workforce section. The workable position is transparency about the measure combined with restraint about the consequences attached to it, because it is the consequence, not the visibility, that drives the search.
What we cannot verify
The two formulations named above are widely cited and also widely argued over in their original disciplines, and neither was written about monitoring software. The examples given are illustrative patterns reported across the industry rather than measurements of any specific organisation. We have not measured how often gaming occurs, and we are sceptical of anyone who claims to have, since a successful gaming strategy is invisible to the system it defeats.
The short version
- A regularity collapses once pressure is put on it for control purposes.
- The correlation was measured in a world where nobody was trying to move it.
- Most routes that move a number are cheaper than doing the work.
- A commercial market in mouse-moving devices is the clearest evidence of the mechanism.
- Pair every metric with a counter-metric, and shorten the distance to the outcome.
- It is the consequence attached, not the visibility, that drives the search.