Amazon’s facial recognition system’s accuracy is being questioned after 28 members of Congress were wrongly matched to criminal mugshots from a database public mugshot photographs.
In a study by the American Civil Liberties Union, all 535 faces of lawmakers in the House and Senate were scanned against the faces of the 25,000 mugshots, using Amazon’s Rekognition application programming interface.
Included in the list of 28 lawmakers who were falsely identified as criminals were Rep. John Lewis, D-Ga., Sens. John Isakson, R-Ga., and Ed Markey, D-Mass. ACLU also noted that a high percentage of those caught up in the false matches were people of color.
[Also read: Two senators demand answers from Amazon’s Jeff Bezos about privacy concerns for ‘Alexa’]
We used Amazon’s facial recognition tool to compare photos of members of Congress to a database of mugshots — we got 28 false matches.
And even though they only make up 20% of Congress, nearly 40% of the false matches in our test were members of color. https://t.co/WdNRWtqZfa
— ACLU (@ACLU) July 26, 2018
“An identification — whether accurate or not — could cost people their freedom or even their lives,” the group said in an accompanying statement. “Congress must take these threats seriously, hit the brakes, and enact a moratorium on law enforcement use of face recognition.”
An Amazon spokesperson told the Verge that the faulty matches could be attributed to poor calibration, and the fact that it was using a confidence threshold of 80 percent — the default setting. Amazon recommends that if law enforcement were to use the technology, it should be operating at least at a 95 percent threshold since the results of a false identification could lead to more significant consequences.
“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 spokesperson said.
