Identifying a face from a photograph · Contested (since 2018)

A study published in PNAS on 29 May 2018 scored a face recognition algorithm at an area under the curve of 0.96 on a hard identity matching test, against a median of 0.93 for professional forensic facial examiners.

The same study found that pairing one examiner with that algorithm gave a median of 1.0, better than pairing two examiners, so the best result on record belongs to neither alone. NIST measured a twentyfold drop in search failures between 2014 and 2018, from four percent to two tenths of a percent, across 127 algorithms from 39 developers. What is settled is scale, since searching a hundred million images is machine work and no person has ever done it. What is not settled is the top of the range, because the 2018 comparison was against the median examiner and not the best one. And recognising a familiar face across fifteen years, a bad angle and a change of weight, with no reference image to hand, is a different act that people perform without effort and that no benchmark measures.