Recognizing someone’s face is hard enough as it is for humans, but even robots can’t get it right 100% of the time. New tests by the American Civil Liberties Union reveal that facial recognition software by Amazon has falsely marked 28 members of Congress as criminals.
The facial ID software “Rekognition” was tested by cross-referencing 25,000 public mugshots with photos of the members of Congress. The test came back with the software’s best guesstimate of 28 members “matched” with the faces of criminals. ACLU explains that the mugshots used included all ages and ethnicities, but that most of the false IDs occurred with people of color… A grim suggestion for the software’s use by police in real investigations.
Amazon is defending its facial recognition software, basically saying the ACLU didn’t tweak the software’s settings enough to reduce the chances of false IDs. They have admitted, however, that 80% accuracy “wouldn’t be appropriate for identifying individuals with a reasonable level of certainty.”
Source: The Verge