The difficulty lies in the features a system can measure
A facial-recognition system does not understand identity the way a person does. It analyzes an image and converts visible patterns—such as the spacing of the eyes, the shape of the nose, and contours around the face—into a numerical representation called an embedding. It then compares that representation with one or more stored faces.
Identical twins inherit nearly the same facial structure, so many of those measured patterns overlap. In effect, the system may receive two faces whose numerical representations sit unusually close together. Small differences can still exist, but they may be harder to detect than the differences between unrelated people.
A close match can matter at the decision threshold
For a one-to-one check, the system compares a face with a specific enrolled identity—for example, when unlocking a device. For a search, it may compare the face with many stored identities. In either case, the system uses a threshold: a similarity score must be high enough to count as a match. If twins’ scores are close, the result can depend on where that threshold is set and on the quality of the images.
Lighting, camera angle, expression, and image resolution can also change the measured face. Those conditions can blur the small distinctions that might help separate twins, although performance varies by system and situation. The practical consequence is that a face match is a comparison result, not proof by itself that the person in front of a camera is the enrolled individual.
