Randomized screening

Why does picking travellers for extra screening at random catch more than screening the ones who look most suspicious?

A screening budget is always too small. Only a fraction of travellers can be pulled aside, and every instinct says to choose well — look at the ticket, the route, the luggage, the manner, and spend the scarce checks on the people who look least ordinary.

And yet security systems deliberately throw dice, flagging travellers by a lottery that ignores everything known about them. That looks like discarding information for nothing. Under what circumstances could deliberately not thinking beat thinking?

Start by asking what the screening is aimed at, because that is the whole answer. A medical screen is aimed at a disease, and the disease does not know which symptom the clinic tests for, nor change its symptoms when it finds out. Airport screening is aimed at a person who chooses what to carry, when to fly, and whom to send.

So imagine you are on the other side. The rule is "young men, travelling alone, one-way ticket, paid cash". You cannot read the rule, but you can probe it: send people through, watch who is pulled aside, and after enough trips its shape is visible. What then? You send someone outside it. The profile has not merely become less effective — it has become a map of where screening is not looking, and you have handed it to the one person capable of using it.

Notice the structure. Any rule applied deterministically can be discovered, and once discovered its catch rate against a competent adversary falls towards zero, while the checks it consumes land entirely on people who — precisely because the rule is known — are no longer the ones being used. A profile is worst exactly against the adversary who took it seriously.

So what would a rule that cannot be discovered look like? It must leave the attacker unable to say where the checks will fall even knowing the whole procedure. Only one family of rules does that, and it is the family involving a random draw. This is why a goalkeeper has no best fixed direction and why inspection regimes visit by surprise: against an opponent who best-responds to whatever you do, unpredictability is not a concession but the strategy.

There is a second, quieter reason, arithmetic rather than game theory. Sample in proportion to how suspicious each traveller looks — the natural "spend the budget where the probability is" rule — and the people at the top of the list are selected again and again, trip after trip, while everyone else is effectively never selected. You are buying repeated confirmation about the same individuals. I recall a formal treatment by William Press, around 2009, showing that such strong proportional profiling performs no better than uniform random sampling for exactly this reason, and that the optimum lies between the two: sampling in proportion to the square root of the suspicion. Treat the precise result as recalled rather than verified; the direction it points is the one the reasoning above already reached.

Now the honest correction, because the question as posed is not quite right. Does random screening literally catch more? Mostly no — and the statistic is near meaningless either way, since the base rate is so low that almost nothing is caught by any method. What randomization changes is how many attempts are made at all. An adversary facing a known profile plans around it; one facing an unknowable draw must accept an irreducible chance of interception every time, and some plans are abandoned. That benefit shows up as things which did not happen, and a regime that deterred perfectly would catch nothing and look like a waste of money.

The refinement connecting the pieces is that adaptivity, not rarity, is what makes randomness pay. Rarity has its own consequence — a screen looking for something present in one traveller in millions produces overwhelmingly false alarms whatever its accuracy, which is the arithmetic worked through in screening-false-positives.md. That effect is present here too but is not the mechanism in question: it applies just as forcefully to a disease, which cannot respond. The fixed point of difference is that a smuggler can watch you and a tumour cannot.

Which gives a clean falsification test. If adaptation is the mechanism, profiling should beat randomization decisively against non-adaptive targets — travellers unknowingly carrying prohibited plant material, or a forgotten bottle in a cabin bag, are selected by observable correlates and nobody is gaming the rule. If profiling wins there while randomization wins against smuggling and attack planning, the account stands. The refuting observation would be random sampling outperforming targeted inspection even for those unwitting cases, which would show something other than adversarial response was doing the work.

Two honesty points. Real systems are not lotteries: they layer watchlists, behavioural detection of doubtful evidential value, and a random component — which is what a mixed strategy looks like in practice. And randomization is favoured for reasons unrelated to any of the above: it is cheap, needs no intelligence input, is auditable, and does not fall foul of laws against selecting people by nationality or ethnicity. Those structural facts explain much of its adoption, and it would be a tidy just-so story to credit the game theory with all of it.

• How inspection regimes in arms control and tax audit set their surprise-visit rates, and whether the same mixture appears.
• Why behavioural detection programmes have proved so hard to validate, and what that implies for layering them under a random component.

Every term the collection defines is gathered in the glossary.