Can Amazon's facial-recognition software tell the difference between a US lawmaker and an arrested felon?

The American Civil Liberties Union doesn't think so. The group recently tested Amazon's controversial Rekognition system and found that it incorrectly matched the photos of 28 US lawmakers with mug shots of people who had been arrested for a crime.

"These results demonstrate why Congress should join the ACLU in calling for a moratorium on law enforcement use of face surveillance," ACLU attorney Jacob Snow wrote in a Thursday blog post.

The ACLU ran the lawmakers' photos against a database of 25,000 mug shots of publicly available arrest images. Amazon's Rekognition system—which anyone can try—is designed to search entire libraries of pictures for facial matches and return "high confidence" results.

That's a big reason why at least a few police departments have been interested in the technology. Amazon's system can potentially help detectives identify criminal suspects caught in security camera footage by matching their faces with existing mug shots. For a human, this might take days, but Amazon's system can find potential matches within minutes.

However, the ACLU is among the groups worried about the dark side of facial-recognition systems; the same technology can also make mistakes. If Amazon's Rekognition is added to body cameras, the system could also pave the way for real-time surveillance and perpetuate wrongful arrests, they claim.

"It's not hard to imagine a police officer getting a 'match' indicating that a person has a previous concealed-weapon arrest, biasing the officer before an encounter even begins," the ACLU's Snow wrote in his blog post. "Or an individual getting a knock on the door from law enforcement, and being questioned or having their home searched, based on a false identification."

For these reasons, the ACLU has been demanding Amazon stop providing its facial-recognition technologies to governments. However, Amazon is pushing back against the ACLU's test-run. The e-commerce giant noted that the photo matching occurred when the ACLU was using technology on its default setting, which shows image matches at an "80 percent confidence" rating or more.

"While 80 percent confidence is an acceptable threshold for photos of hot dogs, chairs, animals, or other social media use cases, it wouldn't be appropriate for identifying individuals with a reasonable level of certainty," the company said. "When using facial recognition for law enforcement activities, we guide customers to set a threshold of at least 95 percent or higher."

The company also played up the positive aspects of its Rekognition product. The technology has already been used to stop crimes involving human trafficking and child exploitation. In addition, the product isn't designed to be a standalone solution to help the police catch criminals; it's merely a tool that officers can use to consider their options.

"We remain excited about how image and video analysis can be a driver for good in the world, including in the public sector and law enforcement," the company said in an email.

The ACLU declined to provide mug shots of the photos the US lawmakers were matched to. "Unfortunately, we're not sharing the mugshots out of respect of the individuals' privacy. The arrestees are not necessarily guilty [of] anything and we want to be mindful of that," the group told PCMag in an email.

However, the ACLU's test results with Rekognition were replicated by Joshua A. Kroll, a computer scientist at the University of California, Berkeley, who used the same database of mug shots. "We were able to get the same 28 false positive matches," he told PCMag in an interview.

Kroll finds that troubling, given that Amazon is marketing the product to law enforcement. But he also takes issue with Amazon's response to the whole matter. Kroll hasn't found any documentation of Amazon telling customers to raise Rekognition's confidence threshold to 95 percent when testing it on human faces. "I suspect there are reasons we don't see any of this," he said. "If you set that confidence value to 95, you'll get too few image matches. You don't get any false positive, but you'll also get no matches at all."