Teams usually file a bad result under one heading: the thing didn't work. That heading covers six published cases where the intervention flipped direction and delivered the opposite of its design goal. Each flip has its own documented mechanism.
Those six mechanisms already appear elsewhere on this site. Spotting which one a live initiative is exposed to before launch turns the flip from a post-mortem surprise into a check someone can run in advance.
The story in four parts
Late fee, bottled-water ban, scare tour, neighbour letter, bounty, and reminder email: each moved the target measure the wrong way.
Crowding out, substitution, a mistaken audience theory, a masked subgroup, a gamed proxy, and a second number nobody tracked.
Daycare centres, a campus shop, juvenile justice, household energy, colonial Hanoi, and charity fundraising all landed on that shape.
Once a team can point to which of the six it faces, the test can run upstream of shipping instead of waiting for the dashboard to go red.
The evidence at a glance
Gneezy and Rustichini's 2000 field study across 10 Israeli daycare centres found a modest lateness fine roughly doubled late arrivals, and the rise held after the fine was withdrawn.
A Campbell Collaboration review of nine randomised and quasi-randomised trials found “Scared Straight”-style programmes raised reoffending by 1 to 28% versus untreated controls.
A Danish charity's field experiment sent a second reminder to a subset of 17,000 past donors: one-time gifts rose, and permanent unsubscribes rose with them.
Six mechanisms share one practical hinge. A clearer causal theory, tighter measurement, or both would have flagged every reversal below before anyone shipped it.
The connection, in short
The causal idea was wrong. Careful delivery cannot rescue a broken theory of how people respond.
Without naming the test as a test, none of these flips get measured closely enough to catch.
Skipping measurement would have left every one of these flips invisible rather than stopped them.
Six cases sit on six different rungs of the same causal ladder.
Begin with the mechanism most people recognise before they have a label for it. A fine can turn a moral obligation into a fee someone can pay off. Once lateness is priced, the guilt that used to keep most parents punctual has nowhere left to attach.
Why did adding a fine for lateness make parents show up later, not sooner?
Staff at ten Israeli daycare centres were tired of parents arriving late to collect children, so they introduced a modest lateness fine. Late pickups did not fall. They roughly doubled, and stayed high after the fine was later withdrawn.
What it teaches: A fine can “crowd out” an existing moral motivation and leave a transactional one in its place. Before the price tag, lateness felt like imposing on the teacher. Once money settled the debt, parents could buy the extra time. The guilt that had kept most of them punctual stayed gone even after the price itself disappeared.
Crowding out swaps one motivation for another. Substitution swaps one behaviour for another, sending the same underlying want to the next-closest option, which can be worse on the exact dimension the policy hoped to improve.
Why did banning bottled water increase the plastic shipped to campus?
In 2013 the University of Vermont banned bottled-water sales on campus, aiming to cut plastic waste and push people toward fountains and reusable bottles. Beverage sales tracked before and after showed sugary-drink purchases rising substantially. Total plastic bottles shipped to campus, now mostly soda and juice rather than water, went up rather than down.
What it teaches: Taking one option out of a choice set leaves the underlying want intact. Demand moves to whatever sits next closest, and that substitute can worsen the metric the policy cared about. A policy aimed at one outcome has to be scored against every outcome it could plausibly move, including the ones it was never designed around.
Crowding out and substitution both assume the audience still wants roughly the same thing, only routed differently. The next mechanism is harder: the intervention can teach the opposite lesson from the one on the slide deck.
Why did showing teenagers prison make them more likely to reoffend, not less?
The premise sounds obvious: take at-risk teenagers through the harsh reality of prison and they will steer clear of crime. The programmes stay popular and still get revived, including on reality TV. A systematic review of nine randomised and quasi-randomised trials found participants were, on average, more likely to reoffend afterwards than a matched comparison group who received no intervention.
What it teaches: A story that sounds right is still only a story until someone tests it. “Scaring” someone off crime is a reasonable-sounding theory of behaviour change, and the randomised evidence points the other way. Funding and re-runs can continue for decades on that story alone, with no proper control-group test ever locking the claim down.
Scared Straight's flip showed up across the whole treated group. The next case buries its flip inside one slice of the population, invisible in the average until that population is actually split.
Why did telling frugal households they used less energy make them use more?
A field experiment sent households a letter comparing their energy use to the neighbourhood average. Among households already above that average, usage dropped. The letter did what it was meant to do there. Among households already below average, usage rose: once they saw they had “room” to relax, they used more energy than before any letter arrived.
What it teaches: Descriptive social norms (see Social Norm) are a real, well-replicated mechanism that can still push a subgroup the wrong way. That shows up when an intervention is only checked on the people it was designed to help, rather than on the full range it will actually reach. The same research programme found the fix: an injunctive signal (a simple smiley or frowny face) stopped the boomerang while keeping the cut for high-usage households.
The energy-bill letter flipped on a subgroup nobody split out in advance. The next case flipped because the reward sat on a measurable stand-in for the goal, while the goal itself went unpaid.
Why did paying for rat tails end up breeding more rats, not fewer?
In 1902, French colonial administrators in Hanoi paid a bounty for every rat tail turned in, hoping to control a plague-linked rat population. Rat-catchers quickly trapped rats, cut off the tails, and released the animals back into the sewers to breed. Health inspectors soon found farms raised specifically to supply tails for profit.
What it teaches: This case is documented colonial-administrative history, reconstructed from municipal archives. It is illustrative history rather than a designed experiment, flagged the same way this site flags any case that is not a controlled study. It remains a genuine, well-sourced example of a perverse incentive. Paying for a proxy (tails) instead of the actual outcome (fewer rats) let people optimise for the reward in a way that made the real problem worse.
Rewarding a proxy is a measurement mistake made before an intervention runs. The last case is a measurement mistake made after it: the tracked number improved, and a different number nobody watched got worse.
Why did a second reminder raise donations and drive donors away for good?
A Danish charity emailed reminders to roughly 17,000 past donors, and a subset also got a second reminder a week later. The extra reminder did its designed job. One-time donations rose substantially. Donors who received it also unsubscribed from future contact at a significantly higher rate. A follow-up test found announcing monthly reminders (versus a single future one) drove even more people to opt out permanently.
What it teaches: A backfire can sit outside the metric being optimised. On short-term donations alone, the reminder looked like a clean win. The real cost shows up in a different number, months later, once the donors who can never be reminded again are counted. Watching the metric you moved leaves every other move unmeasured. This mechanism now has its own Principles entry: Reminder Fatigue.
Six mechanisms, one structural gap. Every case above cleared whatever check its own team was already running. None of them had a check built for the specific way that intervention could reverse, because that mechanism had never been named as a shipping risk.
The six cases above are specific and varied. The four checks below travel: they can run on almost any intervention before launch, whichever of the six it happens to risk.