Skip to content
Final StateThe Base Rate Trap: When the False Alarm Rate Decides
VOL. I  ·  NODE 115▢  ATLAS

THE WATCHMAN'S ALARM

The Base Rate Trap: When the False Alarm Rate Decides

The base-rate trap is rare-event arithmetic: even a low false-positive rate, multiplied across many non-events, can bury the true hits.

RARITY MOVES THE GOALPOSTS

For a rare event, specificity and false positives decide

Rare-event diagram showing a tiny real-event pile beside a huge non-event pile.The figure shows why false positives become binding: a small false-positive rate applies to the enormous non-event pile and can swamp true hits.RARITY MOVESTHE GOALPOSTS1 EVENT9,999 NON-EVENTS1% FALSEPOSITIVE RATESPECIFICITY BINDS
  • A detector with 99% specificity still has a 1% false-positive rate
  • When the event is rare, that rate hits the enormous pile of non-events
  • The mathematics of rarity, not just a bad model

The trap is structural. Once the thing you hunt is rare enough, even a small false-positive rate on the enormous pile of non-events can swamp the true hits.

NINETY-NINE OF A HUNDRED

One in ten thousand: about 100 false alerts for one true hit

False-positive arithmetic showing a one-in-10,000 event can produce about 99 false alerts for every true hit.With a rare base rate, even high recall and 99% specificity leave precision low because the 1% false-positive rate is applied to the much larger set of non-events.BASE-RATE ARITHMETICONE IN 10,000ABOUT 101 ALERTS1 TRUE HIT~100 FALSEPRECISION ABOUT 1%~99% OF ALERTS ARE FALSE
Illustrative arithmetic: 1 event and 9,999 non-events; perfect recall and a 1% false-positive rate yield 1 true hit and about 100 false alerts.

You can catch every true event (high recall) and still have almost every alert be wrong (low precision). For a rare target, recall and precision can come apart violently — and precision is the number that decides whether the system is usable.

  • Recall: of the real events, how many you catch
  • Precision: of your alerts, how many are real
  • A 1% false-positive rate can still leave ~99% of alerts false

THE PERSONALITY SKETCH

People often ignore the base rate

Kahneman and Tversky base-rate neglect diagram with prior mixes ignored after a vivid sketch.Two different engineer-lawyer base mixes point to the same personality sketch, showing how resemblance can dominate the stated prior.STORY VS PRIORSTATED MIX70 ENGINEERS30 LAWYERS30 ENGINEERS70 LAWYERSPERSONALITYSKETCHQUIETMETHODICALSTORY FITSAME GUESSPRIOR IGNORED
Source: Daniel Kahneman & Amos Tversky, “On the Psychology of Prediction,” Psychological Review 80(4), 1973, 237-251. doi:10.1037/h0034747.
  • Kahneman & Tversky, early 1970s: a sketch 'drawn at random' from 100 professionals
  • Told 70 engineers or 70 lawyers — the mix barely moved the guess
  • Judgment ran on resemblance, not on the odds

This is not only a property of machines. The founders of behavioural economics showed the human mind does it too — we reach for how well a case fits the story and drop the prior.

INSENSITIVITY TO PRIOR PROBABILITY

Insensitivity to prior probability of outcomes.

Kahneman & Tversky, "On the Psychology of Prediction," 1973

Their own name for the effect. Later shortened to base-rate neglect — the tendency to underweight the prior in favour of the vivid new detail in front of you.

THE INTRUSION DETECTOR

Where it became operational: intrusion detection

Intrusion-alert queue with many false alerts and one true alert arriving after fatigue forms.The figure makes Axelsson's operational point: rare intrusions can make the false-positive rate the variable that determines whether operators keep listening.ALARM FATIGUEFALSE ALERTSTRUEFATIGUE FORMS FIRSTMUTED
Source: Stefan Axelsson, “The Base-Rate Fallacy and the Difficulty of Intrusion Detection,” ACM TISSEC 3(3), 2000, 186-205. doi:10.1145/357830.357849.
  • Axelsson, 2000: for a rare intrusion, the false-positive rate is the binding variable
  • Flooded operators can suffer alarm fatigue
  • Repeated false alarms can get normalized — then muted

Then the second failure. Through Vaughan's normalization of deviance, a team living with constant false alarms can learn to wave them off — including, one day, the true one.

The trap doesn't just flood you. It teaches you to ignore the truth.

A threshold chooses an operating point on the ROC tradeoff: lowering false positives usually lowers recall too. At the real base rate, choose the point from false-positive cost, miss cost, and required precision.

WHAT SURVIVES THE FLOOD

What gets through: joining, triage, cheap probes

Three rare-event remedies: join independent signals, triage evidence, and run a cheap probe.The diagram conditions conjunction on sufficiently independent errors and marks its false-positive benefit against possible recall loss before triage and probing.SURVIVE THE FLOODJOINSIGNALSAGREETRIAGEEVIDENCEPROBECHEAPERRORS MUST DIFFERFALSE ALERTS DOWNRECALL MAY DROP
  • With sufficiently independent errors, conjunction may lower false positives
  • Requiring agreement can also sacrifice recall
  • Treat the output as triage for a cheap probe, not a conclusion

Conjunction helps only when channels add sufficiently independent evidence and the recall loss is acceptable. Then route the smaller alert set to a real-world verifier.

THE FIRST ENEMY

The first enemy every rare-event watcher meets

  • Rarity floods the signal — this trap
  • Fluency fools the judge — the metacognitive demand
  • Operator test: at this base rate, how many alerts are true?

At the jagged edge, rarity can flood alerts and fluent output can mislead the judge. Ask what share of alerts will be true at the deployed base rate.

Read the transcript

01 · THE WATCHMAN'S ALARM

A watchman keeps a detector with what sounds like a tight spec: it catches the real alarm, and on quiet nights it cries wolf only one percent of the time. That sounds like a tool you would trust with your life. Listen to what it does to him instead. The nights stack up. Alert, alert, alert, and each one, checked, turns out to be nothing. A fox. A gust. A shadow. He runs them all down anyway, because the one time he doesn't is the time it matters. Except it never seems to matter. Slowly the alerts stop meaning danger and start meaning noise. And on the night the wolf finally comes, its alarm looks exactly like the false alarms that came before. His hand is already halfway to the switch.

02 · RARITY MOVES THE GOALPOSTS

Here is why the good detector fails him. Ninety-nine percent specificity means only one false alarm in a hundred quiet cases. That sounds like the whole story, but it is not. When the event you are hunting is rare, the number that decides the system is not a blended accuracy score. It is the false-positive rate at the real base rate. Think about the arithmetic underneath. There is a vast pile of non-events, and a sliver of real ones. Even a tiny false-positive rate, applied to that vast pile, throws off a mountain of false alarms, while the true events, being rare, stay a sliver. The detector can be useful and still bury you if the base rate is low enough. This is not just a bad model. It is the mathematics of rarity, and a summary accuracy figure can hide it.

03 · NINETY-NINE OF A HUNDRED

Let us do it slowly, with numbers you can hold. Say the event strikes once in ten thousand. Give the detector perfect recall and ninety-nine percent specificity. It catches the one real event, and misfires on one percent of the nine thousand nine hundred and ninety-nine non-events. That produces about one hundred false alarms. So the alert pile holds about one hundred and one flags, and just one is real. Roughly ninety-nine percent of the alerts are wrong, and precision is about one percent. Recall says you caught the event. Precision says the operator now has to find it in the noise.

04 · THE PERSONALITY SKETCH

And this is not only a machine problem. It is an old human one. In the early nineteen-seventies, two psychologists, Daniel Kahneman and Amos Tversky, ran a now-famous test. They handed people a short personality sketch, said to be drawn at random from a hundred professionals. One group was told the hundred held seventy engineers and thirty lawyers. Another was told the reverse. That mix, the base rate, should swing a probability-aware guess. It barely moved the needle. People judged by how much the sketch sounded like an engineer or a lawyer, and all but ignored the odds. The founders of behavioural economics had caught the human mind doing a version of what the watchman does.

05 · INSENSITIVITY TO PRIOR PROBABILITY

They gave the effect a careful name: insensitivity to prior probability of outcomes. Say it slowly, because every word earns its place. Insensitivity. Not ignorance; we can state the base rate if asked, we simply do not feel its weight. Prior probability. The odds before any evidence arrives, the seventy in a hundred. Of outcomes. The thing you are trying to predict. Later it got a blunter name: base-rate neglect. The vivid detail in front of you, the story that fits, crowds out the quiet number that should have anchored the whole judgment. And a fluent, confident AI flag can be exactly that kind of vivid detail. So a machine flag does not automatically cure the bias. In the wrong design, it supplies it at scale.

06 · THE INTRUSION DETECTOR

Security researchers ran straight into this wall. In 2000, Stefan Axelsson formalised it for intrusion detection: when the attack is rare, the variable that can bind the whole system is the false-positive rate. If that rate is too high, the operator drowns. But drowning is only the first harm. The second is quieter. Live with false alarms long enough and a team can stop treating them as alarms at all. Diane Vaughan's term normalization of deviance describes the related organizational drift: repeated departures become the background people learn to accept. Applied here, the flood gets muted, and the one real signal can be muted with it.

07 · Advertisement · Bubble AI App Builder

Some ideas do not need another document before they become testable. With Bubble AI, you describe the app you want, and Bubble creates a working starting point: interface, data, and logic you can inspect. From there, you refine visually, connect AI models and services, and turn the first version into something real enough to use.

08 · THE ONE YOU MUTED

A threshold is not simply right or wrong. It selects an operating point on the detector's ROC tradeoff. Raise the threshold and the false-positive rate may fall, but recall usually falls as more true events are missed. Lower it and recall may improve while the alert queue grows. At the deployed base rate, precision follows from that operating point and the event prevalence. Choose the threshold from the cost of false alarms, the cost of misses, capacity to investigate, and required precision. In some systems a threshold can meet those requirements; in others no available point can.

09 · WHAT SURVIVES THE FLOOD

Conjunction can help, but only under conditions. If channels make sufficiently independent errors, requiring several signals to agree may lower the joint false-positive rate. The same requirement can miss true events that appear in only one channel, so recall may fall. Correlated channels can simply repeat the same error and provide little gain. Measure both sides at the deployed base rate. Then use the surviving flags for triage: narrate the evidence, state a counter-argument, and run a cheap real-world probe before acting. The conjunction is a filter with a tradeoff, not a truth machine.

10 · THE FIRST ENEMY

So this is the first enemy of every rare-event check you build. Rarity floods the signal until the truth is a needle in a field of needles. And it does not travel alone. Its partner is fluency, the trouble of judging work in a field you cannot command, where confidence reads as skill. Rarity buries the true signal. Fluency talks you into the false one. Between them, they can make the verifier you were counting on fail without a sound. Which returns us to the one honest reading a machine can offer at the edge. Not a louder alarm. A quieter, rarer word: I don't know. That word is one thing the flood cannot easily fake.

01 / 10 · THE WATCHMAN'S ALARM0:00 / 8:05