In January 2020 the Detroit police arrested Robert Williams on the lawn of his home, in front of his wife and daughters, and held him thirty hours. He had done nothing. An algorithm had matched his driver’s licence photo to a grainy shop-camera still and flagged him. It is the plainest lesson there is in how facial recognition works, and fails.
How facial recognition works, in plain terms, is this: software finds a face in an image, measures its geometry into a long list of numbers, and compares that list to others. It never decides who you are. It reports how similar two faces look, as a score, and a threshold or a person decides what that score means.
That gap — between a similarity score and a claim about identity — is where nearly every failure lives. The technology is often accurate and unequal at the same time, and understanding why means following the numbers rather than the marketing.
From a face to a string of numbers
The first job is detection: finding that there is a face in the frame at all, and where its edges are. This is not recognition. Your phone’s camera drawing a box around a face is doing detection, nothing more.
Next the face is normalised. The image is rotated, scaled and cropped so the eyes and mouth sit in standard positions, then the lighting is evened out. A three-quarter pose is a harder start than a passport photo, and the quality of this step quietly sets a ceiling on everything after it.
Then comes the part people mean when they say a face is “read”. A neural network trained on millions of faces converts the normalised image into a vector of a few hundred numbers. This vector is the template, sometimes called a faceprint — not a picture, and impossible to view as one.
The template is a point in a mathematical space, placed so that two photos of one person land close together and two strangers land far apart. Everything downstream is geometry on those points.
A facial recognition system stores these templates, not photographs, and does all its work on the numbers. That is the whole trick, and also the whole problem.
Matching is a distance, not a yes or no
To compare two faces, the system measures the distance between their templates in that mathematical space and turns it into a similarity score. Close together, high score. Far apart, low score.
Notice the wording: the software does not answer “is this the same person”. It answers “how similar, on a scale”. Turning that score into a decision requires a cut-off — a threshold. Above the line counts as a match; below it does not. Move the line and you change every result the system produces without touching the algorithm at all.
This is why the same system can be described, honestly, as both reliable and dangerous. Set the threshold high and it rarely calls a match, but misses real ones. Set it low and it catches more, at the cost of flagging strangers as each other. The line is a choice, and it is usually made by the buyer, not the builder.
Verification and identification are different jobs
The word “facial recognition” hides two tasks that fail in different ways. Getting them confused is how a tool that works fine on a phone ends up jailing the wrong man.
Verification is one-to-one. You claim to be someone; the system checks your live face against the one template on file. This is your phone letting you in, or a passport e-gate. One comparison, and the base expectation is that the claim is true.
Identification is one-to-many. The system takes an unknown face and searches a database of thousands or millions for the closest matches, returning a ranked list of candidates. Nobody has claimed anything. The software is guessing, from a crowd.
The difference matters because errors scale with comparisons. Suppose a system wrongly matches two strangers just once in a hundred thousand tries — very good odds. Run that face against 1.6 million mugshots and you have made 1.6 million attempts, so a few confident false matches are not bad luck but arithmetic. It is the statistics that also sink most clinical trials.
| Verification (one-to-one) | Identification (one-to-many) | |
|---|---|---|
| Question asked | Are you who you claim to be? | Who, if anyone, is this? |
| Comparisons run | One | Thousands to millions |
| Typical use | Phone login, passport gate | Police database search, watchlist |
| Main failure | Lets an impostor through | Confidently names an innocent stranger |
What NIST found when it tested 189 algorithms
The most authoritative measurement of who these systems fail comes from the National Institute of Standards and Technology, a US government lab that has quietly tested face algorithms for two decades. Its 2019 study is the one worth knowing.
NIST evaluated 189 algorithms from 99 developers against 18.27 million images of 8.49 million people, drawn from State Department, Homeland Security and FBI databases. That scale is why its findings carry weight that a single vendor’s demo never could.
In one-to-one matching, the lab reported higher rates of false positives for Asian and African American faces relative to white faces, with the differential often ranging from a factor of ten to a factor of a hundred depending on the algorithm. Among systems built in the United States, American Indian faces drew the highest false-positive rates of all.
One-to-many searching was worse where it hurts most. Testing against a database of 1.6 million domestic mugshots, NIST found the highest false-positive rates for African American women — precisely the group most exposed to a wrong match becoming a knock at the door.
One finding cuts against easy despair. The gaps were not uniform across the field: some algorithms were far more even-handed than others, and the most equitable systems also ranked among the most accurate overall. Bias here is not a law of physics. It is a property of specific software, some of it fixable.
Why does facial recognition misidentify some faces more?
Three forces stack up here, and they compound one another.
- Training data. A network learns the faces it is shown. If those skew toward one demographic, the model becomes better at telling those faces apart and blurs the ones it saw less often.
- Image capture. Camera exposure and lighting have historically been tuned for lighter skin, so darker faces arrive underexposed, giving the algorithm less signal to work with before it starts.
- The numbers game. When one group’s templates sit slightly closer together on average, a fixed threshold produces more false matches inside that group — a small statistical bias amplified by every large search.
None of this requires a bigoted engineer. It is what happens when a general-purpose tool trained on uneven data is pointed at a task where the cost of error is a person’s liberty. That is the honest limit of the technology, and the reason NIST keeps publishing.
The confidence threshold is a policy dial
Return to the cut-off, because it is where the argument is really settled. Vendors sell a candidate list with a similarity score attached to each name. The buyer decides how high a score has to be before a human acts on it, and whether a match is a lead to investigate or something closer to proof.
Set responsibly, one-to-many results are treated as a pointer — a reason to look, never a reason to arrest. Set carelessly, a ranked guess becomes the case. The Williams arrest was not really an algorithm failing; it was a low bar and a detective who mistook a suggestion for an identification.
Who is using it, and how little is tracked
Facial recognition is already woven into airports, phones and policing, often without its users knowing much about the tool. In 2021 the Government Accountability Office surveyed 42 federal agencies that employ law enforcement officers. Twenty reported owning or using facial recognition systems, and fifteen used systems built by outside companies — yet only one could say which of those systems its staff were searching.
Read that again. Agencies were running searches on commercial databases and could not name the databases. Without that, they cannot assess a system’s accuracy or its demographic error rates, which is the very thing NIST’s work exists to measure.
Compare this with how medical devices are approved before a clinician may rely on one. A face-matching algorithm can be sold to a police department with no equivalent gate, no required accuracy disclosure and no standard for how its output may be used in a case. The technology outran the rulebook, and the rulebook is only now catching up.
This is worth separating from the tracking people fear most. A face search identifies who appears in an image someone already holds; on its own it does not follow you around a city in real time. That is a different capability, much as GPS locates a device without quietly tracking its owner. Conflating the two muddies which risk you actually face.
What to watch as the rules catch up
The Williams case did more than compensate one man. As part of the 2024 settlement, Detroit agreed that police may not arrest anyone on the basis of a facial recognition match alone, and to audit past cases where the technology led to a warrant. Several states are now writing similar limits into law.
Three signals tell you whether a given deployment is serious about accuracy. Does it publish which algorithm it uses and that algorithm’s NIST results? Does it forbid arrest on a match alone, requiring independent evidence? And does it treat a one-to-many hit as a lead rather than a conclusion?
Where those three hold, the same technology that failed Robert Williams becomes a narrow, checkable tool. Where they do not, the score on the screen keeps getting mistaken for a name.
