“AI” is the new “detection time”
Every FOD detection announcement of the last three years leads with artificial intelligence. FODetect classifies objects with it. Skylark Labs has a Runway Monitoring Intelligence Layer that learns across a global network. ADB SAFEGATE sells AI-powered apron detection. Munich is buying autonomous machines.
Not one of them publishes a probability of detection.
This is the same move the industry made in its first decade with scan time. Detection time is easy to measure, easy to state, sounds decisive, and is almost irrelevant to risk. An object on a runway is dangerous in proportion to how long it stays there and how likely it is to be found at all — FOD minutes, quantity multiplied by dwell time — and a system that scans every thirty seconds while missing four objects in ten is worse than one that scans every five minutes and misses one. That is arithmetic, not opinion, and the calculator has been running it since 2010.
AI is now performing the same rhetorical function. It is a claim about the sophistication of the processing, offered in place of a claim about the outcome. And it is a better shield than scan time ever was, because scan time was at least a number.
The honourable exceptions are worth naming, because there are only two in a decade of coverage.
Changi published a tenfold reduction in false alarms — not a vendor, an airport, reporting on its own installation. The Smart Ramp trials published a ten-times comparison against human inspection, which implies a manual detection rate around ten per cent and is the first honest statement of the incumbent baseline anyone has made. Both are real measurements from live operations rather than datasheet claims — though only the second speaks directly to how much gets found. Neither came from a press release about a product launch.
There is a reason the industry avoids this number, and it is not conspiracy. Measuring probability of detection properly requires seeding a live runway with a specified object set, in representative weather, repeatedly, and counting the misses. It is expensive, it is operationally disruptive, and the result is a number that will be lower than the marketing implies and that a competitor will quote against you. Even the FAA’s cost-benefit analysis had to work from detection counts, because detection counts are what exists.
But a regulator that specifies 95 per cent detection of standard targets forces the measurement into existence. That is the mechanism, and it appears to be operating in exactly one country.
To any vendor reading this: publish the number, name the object set, name who measured it. The first one to do it will own the argument for a decade.