Very small probabilities behave badly in the mind

The range where intuition stops working

People handle middling probabilities reasonably well. Something that happens about half the time, or a fifth of the time, corresponds to an experience most of us have had, and the estimate has something to sit on. Below roughly one in a hundred, that grounding disappears, and the numbers start being processed as categories rather than quantities.

The categories are something like: basically never, very unlikely but it happens, and worth worrying about. A one in ten thousand chance and a one in a million chance land in the same category despite differing by a factor of a hundred, which is the same as the difference between an hour and four days. In any other context that difference would be obvious.

Two errors that look opposite

The first is treating a small probability as effectively zero, which is why people skip precautions against rare, severe events. Nothing in ordinary experience contradicts the assumption, because the event has not happened and probably will not this year either. The assumption is confirmed daily right up until it is not.

The second is treating a small probability as much larger than it is, which is why some rare risks attract attention out of all proportion. These are not really opposite errors — they are the same failure to discriminate, resolved differently depending on how vividly the event can be imagined. The deciding factor tends to be how easily an example comes to mind rather than anything about frequency, which is why heavily reported risks feel common and statistically commoner ones do not.

How the numbers are usually presented does not help

A probability expressed as a decimal is nearly unreadable at small magnitudes; a string of zeros conveys nothing except that it is small. Percentages are barely better below one per cent. Frequencies with a fixed reference class work considerably better: out of a hundred thousand people doing this for a year, this many will experience it.

Comparison against something familiar helps more than any amount of restating. A risk expressed alongside another risk of roughly known size gives the number somewhere to sit, which a bare figure never does. The comparison has to be honest about the differences — one risk may be voluntary and the other imposed, one may be spread over a lifetime and the other concentrated in an afternoon — and those differences matter to people for reasons that are not irrational. But an anchored comparison with caveats beats an unanchored number with none.

There is a related trap in how people compare ratios. Presented with two chances expressed with different denominators, people often attend to the numerators — the count of winning cases — and judge a larger numerator as a better chance even when the ratio is worse. Keeping the denominator constant when comparing removes the problem entirely, and it is one of the few reliable fixes in this whole area.

Repetition changes everything

A single exposure to a one in a thousand risk is negligible for most purposes. The same exposure repeated daily for years is not, and this is where small-probability reasoning most often goes wrong in practice. The mind evaluates the instance and ignores the accumulation, because each instance genuinely is safe and the sequence genuinely is not.

The arithmetic is not intuitive but it is not difficult either. The chance of avoiding an event across many independent attempts is the chance of avoiding it once, multiplied by itself once for each attempt, and that product falls faster than people expect. This is the entire logic of routine safety procedures that seem absurdly cautious for any single occasion. They are not calibrated to the occasion. They are calibrated to the thousands of occasions.

When the small number is itself uncertain

Small probabilities are also the ones we know least precisely, which is a point that gets lost. Estimating how often something happens requires observing it happen, and rare events by definition provide few observations, so the estimate rests on a thin record or on a model rather than on a count.

This means that a stated small probability often carries a wide margin around it, and the margin can be larger than the number. Where the consequence is severe, that uncertainty argues for more caution rather than less, because you are not choosing between the stated risk and zero — you are choosing under a distribution whose upper end you cannot see. The reasonable response is to treat precise-sounding small numbers with more suspicion than large ones, not less.