But the systems used by law enforcement agencies don’t always have the latest algorithms; Pinellas’s, for example, was last overhauled in 2014, although the county has been evaluating other, more recent, products. Idemia declined to comment on it.
The gains in quality of the best facial recognition technology in recent years have been astounding. In government tests, facial recognition algorithms compared photos with a database of 1.6 million mug shots. In 2010, the error rate was just under 8 percent in ideal conditions — good lighting and high-resolution, front-facing photos. In 2018, it was 0.3 percent. But in surveillance situations, law enforcement hasn’t been able to count on that level of reliability.
Perhaps the biggest controversy in facial recognition has been its uneven performance with people of different races. The findings of government tests released in December show that the type of facial recognition used in police investigations tends to produce more false positive results when evaluating images of black women. Law enforcement officials in Florida said the technology’s performance was not a sign that it somehow harbored racial prejudice.
Officials in Pinellas and elsewhere also stressed the role of human review. But tests using passport images have shown that human reviewers also have trouble identifying the correct person on a list of similar-looking facial recognition results. In those experiments, passport-system employees chose wrong about half the time.
Poorer-quality images are known to contribute to mismatches, and dim lighting, faces turned at an angle, or minimal disguises such as baseball caps or sunglasses can hamper accuracy.
In China, law enforcement tries to get around this problem by installing intrusive high-definition cameras with bright lights at face level, and by tying facial recognition systems to other technology that scans cellphones in an area. If a face and a phone are detected in the same place, the system becomes more confident in a match, a Times investigation found.
In countries with stronger civil liberties laws, the shortcomings of facial recognition have proved problematic, particularly for systems intended to spot criminals in a crowd. A study of one such program in London, which has an extensive network of CCTV cameras, found that of the 42 matches the tool suggested during tests, only eight were verifiably correct.
Current and former Pinellas County officials said they weren’t surprised. “If you’re going to get into bank robberies and convenience store robberies, no — no, it doesn’t work that well,” said Jim Main, who handled technical aspects of the facial recognition program for the sheriff’s office until he retired in 2014. “You can’t ask, like: ‘Please stop for a second. Let me get your photo.’”
Kitty Bennett contributed research.