Wrong on Purpose

Nothing was wrong. The alarm that woke you was mistaken, and it was working perfectly. Both of those are true at once, and the reason is one of the most useful ideas I know.

The principleEvery detector makes two mistakes

Anything that watches for danger can be wrong in two ways. It can stay silent when something is there, which is a miss. Or it can sound when nothing is there, which is a false alarm. You cannot push both mistakes to zero with the same sensor. All you can do is move the threshold, and moving it trades one mistake for the other. Lower it and you catch more of what is real, at the price of more noise. Raise it and you get quiet, at the price of more misses.

Where the threshold belongs depends on one thing: what each mistake costs. This is the mathematics radar operators worked out when they had to decide whether a blip was an aircraft, and it became signal detection theory. The physician Randolph Nesse applied it to the body and gave it a name that sticks: the smoke detector principle. A smoke detector that goes off when you burn toast is annoying. A smoke detector that sleeps through a fire is a funeral. So we build them jumpy, and we accept that nearly every time one sounds, it is toast.

Your body is that kind of detector. The jolt at a coiled shape on the trail. The drop in your stomach when a message says got a minute? The stove at 3:12. For almost all of history a miss could cost everything and a false alarm cost a few minutes of fear, so the alarm was tuned to fire on very little. The ancestors who checked are the ones who had descendants. Most of what wakes you is not a broken alarm. It is a good alarm doing arithmetic set for a harder world. (I am talking about the ordinary alarm everyone carries, not anxiety disorders, which are a medical matter.)

A good alarm is supposed to be wrong most of the time. The question is who pays to find out.

The hidden termA false alarm costs what the check costs

Look at the formula again. The cost of a false alarm sounds like a fixed thing, but it is not. A false alarm costs exactly what it costs to discover that it was false. For a smoke detector, that is a glance at the toaster. For you, at 3:12, it is getting out of bed, or a phone call you do not want to make, or an hour of lying awake running the possibilities. In a human life, checking and worrying have always been the same act. You could not verify without paying in attention, so every false alarm was billed to you.

Engineers solved this long ago with a second stage. Put a jumpy, cheap detector first, so nothing real slips past, and behind it put a confirmation step that sorts the real from the noise before anyone acts. Medicine screens and then confirms. Alarm companies call the house before they call the fire department. Your life never had a second stage that cost you nothing. Now it can.

1. The alarmJumpy by design. Fires on very little, because a miss is the expensive mistake.
2. The checkYour delegate looks: the stove, the account, the portal, the message. A false alarm here costs nothing.
3. YouHear only what survived the check, with the receipt attached.

This is the turn. An AI delegate with access to the places your alarms point can be that confirmation stage. It can read the burner, look at the account behind the fraud text, open the school portal, pull up the thread behind got a minute? The first alarm can stay as jumpy as it likes, and should. The false alarms do not vanish. They move to the machine, where they cost nothing, and you become the third stage, the one that hears only what is left.

What most people missYou cannot tell the alarm to stand down

Here is the part that is easy to get wrong. Adding the check does not quiet your body on day one, and no amount of telling yourself to relax will do it. What you expect is not something you set. It is inherited from history, and the only thing that retrains it is more history. The alarm in you learned from every night that mattered, and it will only unlearn from evidence.

So the check has to leave something behind. Not don’t worry, I’ve got it, which teaches nothing, but a receipt: burner off since 9:40, reading attached. Each receipt is a small piece of the record. Enough of them and the 3:12 alarm starts to fire less, not because you chose calm, but because the past finally gave it a reason. Reassurance without receipts is just another voice in the dark.

There is a second thing hiding in the formula. The costs belong to someone. Every institution that watches on your behalf tunes its threshold to its own costs, not yours. A bank’s miss is a loss and a lawsuit; its false alarm is a text message, and you are the one who pays for it with a racing pulse in the checkout line. Every recall notice, every unusual-activity alert, every warning light that turns out to be a loose sensor is someone else’s smoke detector installed in your attention. A delegate that inherits its maker’s cost ratio is just one more of those. The delegate worth having is tuned to what a miss and a false alarm cost you.

PredictionsWhat changes when the check is free

1. The most valuable message becomes “checked, nothing.”

   People will choose delegates less for what they do than for what they quietly verify. The confirmation layer is the product. A delegate that never says anything is useless; one that says “checked, nothing, here’s why” every time an alarm would have reached you is the one people keep.

2. Institutional alerts will be addressed to delegates first.

   The fraud text, the recall, the portal notice will be read, verified and answered by your delegate before you see it, and only confirmed problems will get through. Institutions will start writing alerts for machines, structured and verifiable, because the person is no longer the first reader.

3. Delegates will be judged on their misses, not their noise.

   Quiet is easy. The failure that matters is the silent miss. Trust will come from a visible record of what the delegate checked and what it let through, the way you trust a smoke detector because you have heard it test itself.

4. The body will recalibrate slowly, and unevenly.

   People whose delegates show their work will start sleeping through the 3:12 stove within months. People who are only told not to worry will not. Receipts retrain the alarm; reassurance does not.

5. “Who set this threshold?” becomes a household question.

   When the first stage moves out of the body, the felt risk moves with it, from the stove to the setting. Asking who tuned the delegate, and to whose costs, will become as ordinary as asking who has a key to the house.

Where I standThe same move, one layer down

In The Last Line Is You I wrote about telling my AI “you decide,” because I had become the queue. This is the same move one layer down. Deciding was the queue. Checking is the alarm. I do not want fewer alarms. I want them answered before they reach me, with the receipt attached, so that the part of me that wakes at 3:12 can learn, slowly and from the record, that it may stand down. It is the same mechanism I described in Permission to Forget for unfinished tasks, and the same bargain behind The Zero-Attention Economy: a machine earns the right to be ignored by being checkable.

The claim ledgerHow this could be wrong

Already true

The arithmetic is settled: when a miss costs far more than a false alarm, the right detector is jumpy and mostly wrong. Institutions already set their thresholds by their own liability. Delegates can already read messages, look up accounts and query devices.

What has to happen

Delegates need permission to look where the alarms point, and they must log each check as a receipt a person can inspect. Their thresholds have to be set by the person’s costs, not the maker’s, and the person has to be able to see that setting.

Where I am probably wrong

If the body only trusts its own checking, machine receipts will not retrain it, the second stage will add a step without adding rest, and people will end up checking the checker. If a few years of good delegates leave the 3:12 stove getting people out of bed as often as it does now, I was wrong.

Back to the kitchen. The alarm that woke you will still be wrong most nights. It is supposed to be. A detector that never cried wolf would be a detector that slept through the wolf. The only thing that changes is who gets up to find out.

Not you.

BackgroundSources and related posts

• Randolph M. Nesse, “The Smoke Detector Principle,” Annals of the New York Academy of Sciences, 2001.
• Randolph M. Nesse, “Natural selection and the regulation of defenses: A signal detection analysis of the smoke detector principle,” Evolution and Human Behavior, 2005.
• David M. Green and John A. Swets, Signal Detection Theory and Psychophysics, 1966.
• Martie G. Haselton and David M. Buss, “Error Management Theory,” Journal of Personality and Social Psychology, 2000.
• Surprise Has a Noise Floor
• The Insured Mind