face recognition
/ fays rek-ug-NISH-un /
Face recognition decides whose face this is. The usual pipeline has three steps: first detect that there is a face and where (face detection), then convert that face into a compact list of numbers — a faceprint or embedding — that captures its distinctive geometry, and finally compare that faceprint to others. Two faceprints that sit close together in this number-space are judged to be the same person. It is essentially measuring how similar two faces are, expressed as a distance.
Two jobs hide under one name, and the difference matters. Verification asks "is this the same person as that?" — a one-to-one check, like your phone confirming it is you. Identification asks "who, out of these thousands, is this?" — a one-to-many search, like scanning a crowd against a watchlist. The second is far harder and far riskier: searching a huge database means even a tiny per-comparison error rate produces many false matches in absolute terms.
This is the most ethically loaded topic in this field, and honesty demands plain limits. Accuracy varies sharply with image quality, angle, lighting, age, and crucially with demographics — many systems have measurably higher error rates for women and for people with darker skin, because of skewed training data. A confident match is a statistical similarity, not proof. The consequences of error — a wrongful arrest, a denied entry — fall on real people, which is why face recognition for surveillance and policing is contested, regulated, and banned in some places. The technology working is not the same as it being right to deploy.
Phone unlock is verification: it stores one faceprint of you and checks each unlock attempt against just that — one-to-one, low stakes, and you can fall back to a passcode. A police system scanning a stadium crowd against a 10,000-name watchlist is identification: even a 0.1% false-match rate could flag dozens of innocent people in a single sweep.
Verification (one-to-one) is easy and low-risk; identification (one-to-many) multiplies rare errors into real harm.
Documented bias is the headline caveat: many face recognition systems are markedly less accurate for women and darker-skinned people, so a "match" can be both confident and wrong, with serious consequences. Vendor accuracy figures are measured on favorable test sets and rarely hold up under real surveillance conditions.