Wired published an investigation into a new AI tool being sold to police forces. The reporting is theirs. What follows is the part that is knowable independently: how automatic number plate recognition works, and why a camera that reads plates turns into something rather larger than a camera that reads plates.
The reading is the easy part
An ANPR camera does four things in sequence, many times a second.
- Find the plate. A detector locates rectangular high-contrast regions of roughly the right proportions.
- Correct the geometry. The plate is almost never square to the lens, so the image is transformed back to a flat rectangle.
- Segment and read. Characters are isolated and classified. This is easier than general text recognition because plate fonts are standardised by law.
- Score the confidence. The system knows the plate format for the region and rejects readings that cannot be valid.
Most systems use infrared illumination and a filter, which is why plates appear brightly lit at night. Number plates in most countries are retroreflective by regulation, so they bounce infrared straight back, and the rest of the scene goes dark. It also makes the reading independent of headlights and weather.
Accuracy on a clean plate in reasonable conditions is very high. Errors cluster on predictable confusions: 0 and O, 8 and B, 5 and S, and on plates that are dirty, damaged or deliberately obscured.
The database is the actual system
Reading a plate is unremarkable. What changes the character of the technology is that each read is stored with a timestamp and a location.
One camera gives you a list of vehicles that passed a point. A network of cameras gives you movement: where a vehicle went, when, how often, and by what route. Retain that for months and you have a searchable history of where a car has been.
That is a qualitatively different thing from a camera, and it is created not by better recognition but by retention. The privacy question with ANPR is almost always about how long records are kept and who may query them, rather than about the reading itself.
Vehicle fingerprinting
Newer systems do not rely on the plate alone. They classify make, model, colour, body type, roof racks, damage, bumper stickers and other distinguishing features.
The practical effect is that changing or obscuring a plate no longer defeats tracking, because the vehicle is identifiable by its combination of visual attributes. It also means a search can be run for a description rather than a registration, which widens who can be caught by a query considerably.
Where it gets contentious
- Everyone is recorded, not suspects. ANPR is indiscriminate by design; the filtering happens after collection.
- Retention varies enormously between jurisdictions and operators, from days to years, and is frequently set by policy rather than statute.
- Private networks exist. Car parks, retail parks, petrol stations and residential schemes run their own, and data-sharing arrangements between private and public systems are often opaque.
- Errors have real consequences. A misread plate matching a wanted vehicle can produce an armed stop of the wrong person, which has happened.
- Pattern inference. Repeated visits to a clinic, a place of worship or a lawyer’s office are sensitive facts that emerge from otherwise mundane location records.
What you can actually find out
In the UK, the EU and several other jurisdictions, ANPR records about an identifiable vehicle you own can be the subject of a subject access request, and operators must say what they hold, why, and for how long.
Private operators are covered too. A car park that issues charges from ANPR is processing personal data and owes you the same disclosure.
Obscuring a plate is an offence essentially everywhere and is not a route worth considering. The productive lever is the retention and access policy, which is set locally and is usually published, and which is where any meaningful change happens.
Join the discussion