Part I · Chapter 2 · Machinery · lesson 3
People respond to incentives, including the ones you did not intend
11 min
Part I · Chapter 2 · Machinery · lesson 3
11 min
A hospital is judged on how long patients wait in the emergency department. The target is four hours from arrival to a decision, and it is measured, published and taken seriously.
Waiting times fall. The target is met. Ambulances begin queueing in the car park, engines running, patients inside, because the clock starts when the patient comes through the door.
Nobody decided to do this. No meeting was held. A number was chosen, a hundred separate people each made a locally sensible decision about how to hit it, and the queue moved outside where the measurement could not see it.
A rule is never read the way it was written. It is read as a puzzle: what is the cheapest thing I can do that satisfies this?
The answer is often the thing intended. Frequently it is something adjacent that satisfies the letter of the rule and none of its purpose. The gap between the two is not dishonesty, and treating it as dishonesty is the mistake that keeps the gap open — because the people finding the cheapest route are behaving exactly as designed, and the design is the thing that has to change.
The thing anyone actually wants is usually not measurable. What a hospital wants is patients treated well and quickly. What it can count is minutes between two timestamps.
So a proxy gets chosen, because a proxy can be counted. And a proxy is by definition not the thing — it is something that correlates with the thing under current conditions. The moment the proxy is attached to consequences, people begin optimising it directly, and it stops correlating with the thing, because the correlation depended on nobody trying.
This has a name and the name is worth knowing, because it turns up everywhere: once a measure becomes a target, it stops being a good measure.
Predict
When a rule lands, the reactions arrive in a predictable sequence, and it is useful to be able to name which one you are looking at.
Compliance. People do what was intended. This is the response the designer modelled and often the only one they modelled.
Substitution. People find a different way to satisfy the rule, usually cheaper, sometimes better, frequently orthogonal to the point. The ambulances in the car park are substitution.
Exit. People leave the situation the rule governs. This is the response nobody counts, because the people who left are no longer in the data. A rule that makes something onerous does not produce a record of everyone who declined to do it.
Exit is the most consistently underestimated of the three, because it is the only one that is invisible by construction.
Not the same thing
The outcome someone actually wants, stated in plain words.
A countable quantity chosen because it moves with the goal, under conditions where nobody is gaming it.
Which is which? Put each one on a side.
A school's average exam result across all entered pupils
Pupils who can still do the material in three years' time
Reported crime figures for a district
None of this argues against measuring things. An organisation with no measures is not neutral, it is unaccountable. The useful move is to assume the gaming rather than to be surprised by it, and to ask three questions before the rule ships.
What is the cheapest way to satisfy this without doing the thing? If you cannot think of one, you have not thought long enough. Someone with more at stake will think of it inside a week.
Who is measuring the number, and do they benefit from it? A measure collected by the people it judges is a measure of their willingness to report.
What does the rule do to the people who leave? If the cost of compliance falls hardest on the participants you most wanted to keep, the rule will improve the average by removing them from it.
Check
Almost everything above works identically when the incentive is status, safety or not being shouted at, and those are frequently stronger than the payment.
A rule that makes a manager personally accountable for a rare bad outcome produces extreme caution about that outcome and indifference to every common one, because the common ones are nobody's name on a document. A team told that raising a problem is welcomed, in an organisation where the last three people who did it were reorganised, is responding to the incentive accurately.
The general test is the same: work out what the cheapest satisfying behaviour is, in the currency that actually matters to the person, and expect that.