On a Tuesday in March a man opens his banking app, sees that the money he has put aside every month for two years is now worth less than he paid in, and cancels the monthly transfer. Nothing about his reason for starting it has changed. He was saving for his daughter’s education, which was fifteen years away and still is. What changed is a number on a screen and how he feels about it, and the cancel button happened to be nearest to hand.
The usual explanation for why people undo in a bad week what they decided carefully in a good one is that they panic: they lack discipline, or need educating about staying the course. That is not wrong exactly. It names the behaviour and then stops, which is not the same as explaining it.
So here is the mechanism as best I can work it out, starting with the part I am confident about: Daniel Kahneman and Amos Tversky’s work on how people evaluate outcomes. Their finding was not that losses hurt, which everybody knows. It was that we judge a result not by where it leaves us but by how it compares to a reference point — some particular number we happen to be measuring from. Change that number and the same situation flips from a gain to a loss without anything in the world moving.
Watch that happen to the man in March. In January he was measuring against his goal: fifteen years out, roughly on track, every transfer a step forward. In March he is measuring against the highest number the account ever showed him. Against the goal nothing has happened; a fall this year barely registers across fifteen. Against last month’s balance he is losing, and worse, losing repeatedly — each month the transfer goes out and the total comes back smaller. The standing instruction that was progress in January has become the thing feeding the loss. He has not changed his mind; he has changed what he measures from, and that was enough to reverse the decision.
Which brings me to the part I find interesting as somebody who builds software that moves money. In January the question he was answering was what do I want in fifteen years. In March it is how do I make this stop. Those are different questions with genuinely different right answers, and both get answered with the same tap.
An instruction carries no record of the question that produced it. The app receives a cancellation, and a cancellation is a cancellation, whether it came from a spreadsheet or a bad morning. Systems that move money are built to be fast and not to argue, mostly rightly, since the alternative is a product that second-guesses people about their own money. But that speed is a bet that every instruction deserves it equally.
If you want to design around that rather than against it, the move is to stop trying to fix the frightened person and let the calm one leave a message. There is a real precedent: the retirement scheme Richard Thaler and Shlomo Benartzi designed, where employees agree in advance that part of each future pay rise goes into saving. People sign up readily, and it works because it never asks anyone to accept a cut today. The same shape applies here — the sentence you wrote in January about why you started, shown back to you in March. Shown, and then the cancellation goes through anyway.
Now the strongest objection, which I cannot fully answer. The panic and the real emergency arrive in the same week, by the same cause. The downturn that frightens him may be the one that cost him his job, and stopping the transfer may be exactly right. From inside the app the two are identical: same person, same Tuesday, same tap. Any design that slows the frightened instruction slows the necessary one by exactly as much, and anyone claiming to tell them apart is claiming knowledge of a stranger’s life that a screen does not have.
That is why I would keep such a design to showing rather than blocking, and why I no longer think the interesting work happens at the tap. It happens months earlier.
Except that something is changing, and this next part is reasoning rather than evidence. A model can now see the moment coming. Eleven app opens in a day, the market down, the pattern matching thousands who cancelled within the hour: an ordinary prediction problem, and it will get solved. The trouble is that the signal is indifferent to its use. Knowing somebody’s judgement is at its worst right now is valuable to whoever wants to protect them and equally valuable to whoever wants to sell them something.
So what I now think is this. The reversal is not weakness and not ignorance. It is a change of question that the software cannot see, because an instruction arrives stripped of its reason, and explaining compound growth to somebody does not fix a problem about what they are measuring from. The honest design response is modest: get the calm version of a person to write something down, show it to the later one, and let them overrule it without penalty. The part nobody has settled is who is allowed to act on the fact that the moment has been spotted at all.