Performance varies with image quality, lighting, camera angle, database design and demographic groups. Even a high percentage can produce many false matches when millions of faces are searched. Human review and strict safeguards are essential.
What we found
A system converts facial features into numerical patterns and compares them with stored examples. Verification—checking whether one person matches one account—is different from searching a huge database. Errors can have serious consequences, especially in policing or access decisions. Accuracy claims must therefore include the test conditions, error rates and affected population. The Ridbusters then compare the wording with the original sources, look for the conditions of each observation and check whether another explanation could fit. That turns a surprising headline into a trail a young investigator can follow: claim, test, result and remaining uncertainty. Performance varies with image quality, lighting, camera angle, database design and demographic groups.