When the ACLU recently ran photos of every member of Congress against Amazon's Rekognition database of random mug shots, 28 House members came back as a match. Basically, they were identified as criminals. And a disproportionate amount were people of color.

"Amazon says that the default confidence threshold for facial recognition API—that is, how confident you can be in having a match—is set at 80 percent," Buzzfeed reporter Davey Alba told WNYC. "So they said after this whole blow-up that they recommend clients use no less than 99 percent confidence levels for law enforcement matches."

Back in May, the ACLU released a report noting that Amazon had been pitching its facial recognition software as a tool for law enforcement agencies. Police in Orlando, Florida, and Washington County, Oregon, have been using the program since 2017. 

Now, with evidence that the software disproportionately singles out people of color, will developers be able to address this bias as they refine the technology? 

"Amazon could refine the way the tech readout works," Alba said. "If you're just getting a readout saying that it's 99 percent confident that it's a match, that's troublesome. Maybe we need to rethink the way that readout works." 

Alba spoke with WNYC's Kerry Nolan.