Face recognition and the base rate

Two different tasks

Almost every argument about face recognition confuses two tasks with very different difficulty.

Verification, one to one. Is this face the face on this passport, or the one enrolled on this phone? A single comparison. Modern systems do this extremely well.

Identification, one to many. Who is this face, out of a database of two million? Millions of comparisons, and the system returns the best matches.

The second is not a harder version of the first. It is a different problem, and the reason is the base rate.

The arithmetic

Suppose an identification system compares a face against a database of one million and has a false match rate of one in a million per comparison — a demanding specification. Run one query against a million records and you expect roughly one false match, in addition to a true match if the person is present.

Now put the system on a camera in a station with 50,000 faces passing per day. That is 50,000 queries, each against a million records: fifty billion comparisons a day. At one in a million per comparison you would be generating false matches continuously. Real deployments manage this by setting the threshold very high, which lowers false matches and raises missed detections, and the trade cannot be escaped.

Early UK police trials of live facial recognition produced very high proportions of false alerts among total alerts, and later deployments with better systems and stricter thresholds report far lower rates. Both are true and both matter: the technology improved substantially, and the underlying arithmetic — rare targets in a large stream means most alerts are false unless the threshold is severe — has not gone anywhere.

Unequal accuracy

The US National Institute of Standards and Technology tested 189 algorithms in 2019 and found demographic differentials in most of them: in one-to-one matching, false positive rates were often ten to a hundred times higher for West and East African and East Asian faces than for Eastern European faces, and higher for women, children and the elderly. Some algorithms showed much smaller differentials than others, which is the useful finding — this is a property of a particular system, not of the technology as a category, and it can be tested for.

An ACLU test in 2018 running a commercial identification service against members of the US Congress returned 28 false matches with mugshots, disproportionately of Black members. The vendor argued that the default confidence threshold used was inappropriate for law enforcement — which was a fair technical point and also the whole problem: the default was what a customer would use.

What happens downstream

A face match is a lead. In practice it has repeatedly become a conclusion.

Robert Williams was arrested in Detroit in January 2020 in front of his family, after a face recognition search on poor-quality shop footage produced his old driving licence photograph and an investigator built a photo line-up around it. He was held for thirty hours. In February 2023 Porcha Woodruff, eight months pregnant, was arrested in the same city on a carjacking charge following another face recognition match, and released after eleven hours. In both cases ordinary investigative steps that would have excluded them were not taken. Detroit later agreed to significant restrictions on how such matches may be used.

The pattern is the automation bias of the first module in its most consequential form: the system produced a candidate, and the humans downstream treated a candidate as an identification.

Where the law has landed

The EU AI Act prohibits untargeted scraping of facial images to build recognition databases, and restricts real-time remote biometric identification in publicly accessible spaces for law enforcement to narrowly defined, authorised situations. Several US cities have banned municipal use; some later reversed. India has no dedicated statute, and police systems have been deployed with limited public documentation.

On the private side, the strongest instrument has been Illinois' Biometric Information Privacy Act, which requires consent before collecting biometric identifiers and allows individuals to sue — and has produced settlements in the hundreds of millions.

What to take from this

Three things. A match is a probability, not a name. Accuracy figures quoted for verification do not transfer to identification, and any vendor blurring the two is telling you something. And the deciding question in any deployment is not the accuracy number but what a human is permitted to do with an alert — because that is where a statistical suggestion becomes an arrest.