A new study examining a face surveillance program implemented by the South Wales police finds the system is wrong 50 percent of the time. The Financial Times reports:

The facial recognition technology used by South Wales Police was developed by Japanese company NEC and has two functions: a live camera feed of faces matched against a pre-determined watchlist, known as “locate”, and a comparison of still images of unidentified people against the police custody database, called “identify”. The “locate” mode depends on the quality of cameras following moving targets.

According to researchers, this yielded “true positive” results in 26 per cent of cases and “false positives” in 50 per cent.

“Identify” yielded correct matches in 22 per cent of cases and incorrect matches in 9 per cent. However, more than two-thirds of submissions made by police officers in the identify mode were of images that were not of high enough quality to enable a match.

The study cites poor camera quality, low light, and people wearing things like hats and sunglasses as obstacles to the system’s accuracy. Apparently jogging past the cameras is enough to confuse the facial recognition system.

The ACLU has called for government agencies in the United States to decline to adopt face surveillance technologies, citing both their inaccuracy and the uniquely dangerous character of the surveillance they enable. In July, the ACLU of California used Amazon’s face surveillance product, Rekognition, to scan the faces of members of Congress against a mug shot database. The results? Amazon’s technology misidentified 28 members of Congress in the mug shot database, a disproportionate number of them members of color.