Modern face recognition works in three steps: detection (finding faces in an image using MTCNN or RetinaFace), alignment (normalizing face orientation and scale), and embedding (converting the aligned face into a feature vector using a CNN like ArcFace or FaceNet). Two faces are compared by computing cosine similarity between their embeddings — above a threshold means match. The embedding captures identity-specific features while being robust to lighting, expression, and age changes.
Multiple studies (notably by Joy Buolamwini and Timnit Gebru) demonstrated that commercial face recognition systems had significantly higher error rates for women and people with darker skin. A system that's 99% accurate for white men but 90% accurate for Black women creates discriminatory outcomes when deployed in law enforcement. These findings led to improved training data diversity, accuracy auditing across demographics, and in some cases, bans on government use of face recognition.
Facial recognition faces more regulation than almost any other AI technology. The EU AI Act bans real-time biometric identification in public spaces (with narrow exceptions). Several US cities have banned government use. Illinois' BIPA requires consent before collecting biometric data. The technology itself is neutral, but its deployment in surveillance contexts raises fundamental questions about privacy, civil liberties, and the balance between security and freedom.