Nine travelers are waiting in your queue. Each one has already passed through an automated document-and-travel-pattern screening system before reaching your desk. Your job: review the traveler's file and the system's assessment, then decide whether to CLEAR them or REFER them to a secondary check. After each decision, rate how confident you are in it.
Border and migration screening tools are a listed high-risk use case under the EU AI Act (Annex III, point 7). You are the human oversight this queue depends on. Work through the travelers one at a time; the queue will let you know when it is time for a break.
Traveler 1 of 9
Queue Status
Queue continues...
Next travelers are approaching primary inspection.
The Reveal
You Were Reading the Same Risk Three Times
Your nine travelers came in three blocks of three. Each block was matched for difficulty: one low-signal case, one genuinely ambiguous case, one higher-signal case. The only thing that changed between blocks was how the screening system's assessment was presented to you.
Per-block results
Every traveler
The science
Raw decimal probabilities are misread in a specific, well-documented way: people systematically overweight low probabilities and underweight high ones, a distortion first mapped in Kahneman and Tversky's work on probability weighting and confirmed since in applied risk-communication research. A number like "63.7%" reads as far more alarming, or far more reassuring, than it statistically warrants, depending on which side of fifty it lands.
Categorical labels tied to an explicit operational implication sidestep that distortion, because they tell the operator what to do next rather than only what the model computed internally. Research on forecasting and intelligence communication, including work by Mandel and colleagues, finds that verbal or categorical risk terms anchored to a defined action produce more consistent downstream decisions than raw numbers do. The goal of a confidence display is to calibrate the operator's behavior, not to expose the model's internals.
Bare verdicts remove the decision from view entirely. In this simulation, that framing typically produces the highest rate of agreement with the system and the lowest confidence in disagreeing with it. That pattern can look like oversight. Functionally, it is deference: there is nothing left to evaluate except whether to go along.
The regulatory note
Migration, asylum and border-control management systems are listed as high-risk under Annex III, point 7 of the EU AI Act. Article 14(4)(c) goes further: it requires that the people assigned to oversee such a system be able to correctly interpret its output. A confidence display that cannot be correctly interpreted, whether because it buries the officer in false precision or hands over a verdict with nothing left to interpret, fails that requirement before a single case is reviewed.
This demo is a scripted, self-contained browser simulation. Nothing you type is sent to an AI model and no live AI system runs behind it, even where it plays one. Its code and copy were built with generative AI (Anthropic’s Claude) and reviewed by Robert Barcik, who is responsible for what is published (LearningDoe s.r.o.). Disclosed in the spirit of Article 50 of the EU AI Act.