HILLSBORO, Ore. — When workers at an Ace Hardware here reported that a woman had walked out of the store with an $11.99 tank of welding gas that she hadn’t paid for in her tote bag, an elaborate high-tech crime-fighting operation sprang into action.
A Washington County sheriff’s detective, working with the agency’s Special Investigations Unit, ran the store’s surveillance footage through an internal facial-recognition program built by Amazon, revealing a possible match.
That woman’s license plate was flagged and, three months later, a narcotics officer in an unmarked SUV saw it and radioed other patrol deputies to stop her. A deputy clapped a pair of handcuffs around her wrists, an arrest report states. She said she’d needed the gas to fix her car.
Deputies in this corner of western Oregon outside ultraliberal Portland used to track down criminals the old-fashioned way, faxing caught-on-camera images of a suspect around the office in hope that someone might recognize the face.
Then, in late 2017, the Washington County Sheriff’s Office became the first law-enforcement agency in the country known to use Amazon’s artificial-intelligence tool Rekognition, transforming this thicket of forests and suburbs into a public testing ground for a new wave of experimental police surveillance techniques.
Almost overnight, deputies saw their investigative powers supercharged, allowing them to scan for matches of a suspect’s face across more than 300,000 mug shots taken at the county jail since 2001. A grainy picture of someone’s face — captured by a security camera, a social-media account or a deputy’s smartphone — can quickly become a link to their identity, including their name, family and address. More than 1,000 facial-recognition searches were logged last year, said deputies, who sometimes used the results to find a suspect’s Facebook page or visit their home.
But Washington County also became ground zero for a high-stakes battle over the unregulated growth of policing by algorithm. Defense attorneys, artificial-intelligence researchers and civil-rights experts argue that the technology could lead to the wrongful arrest of innocent people who bear only a resemblance to a video image. Rekognition’s accuracy is also hotly disputed, and some experts worry that a case of mistaken identity by armed deputies could have dangerous implications, threatening privacy and people’s lives.
Some police agencies have in recent years run facial-recognition searches against state or FBI databases using systems built by contractors such as Cognitec, IDEMIA and NEC. But the rollout by Amazon has marked perhaps the biggest step in making the controversial face-scanning technology mainstream.
Rekognition is easy to activate, requires no major technical infrastructure and is offered to virtually anyone at bargain-barrel prices. Washington County spent about $700 to upload its first big haul of photos, and now, for all its searches, it pays about $7 a month.
It’s impossible to tell, though, just how accurate or effective the technology has been during its first 18 months of real-world tests. Deputies don’t have to note in arrest reports when a facial-recognition search was used, and the exact number of times it has resulted in an arrest is unclear. Sheriff’s officials said the software has led to dozens of arrests for theft, violence or other crimes, but a public-records request turned up nine case reports in which facial recognition was mentioned.
“Just like any of our investigative techniques, we don’t tell people how we catch them,” said Robert Rookhuyzen, a detective on the agency’s major crimes team who said he has run “several dozen” searches and found it helpful about 75% of the time. “We want them to keep guessing.”
Sheriff’s officials say face scans don’t always mark the end of the investigation: Deputies must still establish probable cause or find evidence before charging a suspect with a crime. But the Sheriff’s Office sets its own rules for facial-recognition use and allows deputies to use the tool to identify bodies, unconscious suspects and people who refused to give their name.
The search tool’s imperfect results raise the risk of an innocent person being flagged and arrested, especially in cases of the scanned images being blurred, low-quality or partially concealed. Deputies are also allowed to run artist sketches through the search, an unusual use that AI experts said could more often lead to a false match.
Amazon’s guidelines for law enforcement say officials should use Rekognition’s results only when the system is 99% confident in a match. But deputies here are not shown that search-confidence measurement when they use the tool. Instead, they are given five possible matches for every search, even if the system’s certainty in a match is far lower.
After fielding questions from The Washington Post, Amazon added language to those guidelines, stating that officers should manually review all matches before detaining a suspect and that the search “shouldn’t be used as the sole determinant for taking action.”
The relationship between Amazon and Oregon’s third-largest law-enforcement agency is mutually beneficial: The Sheriff’s Office is helping to refine the system, which Amazon hopes to sell across the country. But Amazon’s push into law-enforcement sales has alarmed some legal advocates who say the system poses too many risks to civil liberties. (Amazon founder and CEO Jeff Bezos owns The Post.)
“The government is incredibly powerful, and they bring a lot to bear against an individual citizen in a case,” said Mary Bruington, the director of the Washington County Public Defender’s Office, which represents defendants who can’t afford an attorney. “You couple that with Amazon? That’s a powerful partnership.”
Matt Wood, the general manager of artificial intelligence for the company’s cloud-computing division, Amazon Web Services, said in a statement that Rekognition is just “another input among many other leads for a 100 percent human-driven investigation.”
Still, the company faces criticism on many fronts: Top AI researchers, members of Congress and civil-rights groups — as well as some of Amazon’s own investors and employees — have urged the company to stop providing the technology to law enforcement, pointing to studies that have found that the system is less accurate with dark-skinned faces. Amazon has disputed that research.
Some of Amazon’s rivals have spurned similar contracts. Microsoft President Brad Smith said in April that the company had recently declined to provide its facial-recognition software to a California law-enforcement agency that wanted to run a face scan anytime its officers pulled someone over, but that it had approved a deal putting the technology in a U.S. prison. Microsoft declined to provide details.
Amazon investors will vote in May on a proposal, backed by a group of activist shareholders, that would prevent the company from selling Rekognition to government agencies unless the company’s board determines that it doesn’t pose a risk to human rights.
The Sheriff’s Office allowed Post journalists to spend two days in March in its squad cars, detective’s offices and county jail, observing how deputies have folded the technology into their daily caseload. Most of those interviewed said the software had saved them time, boosted their arrest numbers and helped them process the growing glut of visual evidence. To date, no legal challenge has been made to an arrest on the grounds that the photo match was mistaken, both deputies and public defenders said.
But lawyers in Oregon said the technology should not be, as many see it, an imminent step forward for the future of policing, and they frame the system not as a technical milestone but a moral one: Is it OK to nab more bad guys if more good guys might get arrested, too?
“People love to always say, ‘Hey, if it’s catching bad people, great, who cares,’ ” said Joshua Crowther, a chief deputy defender in Oregon, “until they’re on the other end.”
‘Indistinguishable from magic’
When Amazon revealed Rekognition in 2016, the company called it a breakthrough for a potent style of deep-learning artificial intelligence that showed results “indistinguishable from magic.” In a blog post illustrated with a photo of an executive’s dog, the company offered some general ideas for how people could begin using it, including for security checkpoints or billboards wired to gather data from a viewer’s face.
The unveiling caught the eye of Chris Adzima, a former eBay programmer who had been hired at the Washington County Sheriff’s Office to work on an iPhone app that deputies use to track inmates’ behavior. His agency had hundreds of thousands of facial photos already online and no real way to analyze them. Using Amazon’s AI, he got a system up and running in less than three weeks.
“They didn’t really have a firm idea of any type of use cases in the real world, but they knew that they had a powerful tool that they created,” said Adzima, a senior information systems analyst who works in a small cubicle at the sheriff’s headquarters. “So, you know, I just started using it.”
Deputies immediately began folding facial searches into their daily beat policing, and Adzima built a bare-bones internal website that let them search from their patrol cars. He dropped the search-confidence percentages and designed the system to return five results, every time: When the system returned zero results, he said, deputies wondered whether they’d messed something up.
To spice it up, he also added an unnecessary purple “scanning” animation whenever a deputy uploaded a photo — a touch he said was inspired by cop shows like “CSI.”
As he started flooding Amazon’s servers with image data, account executives there took notice, he said, and some voiced their surprise and excitement that he was using it for police work. In one 2017 email first revealed last year as part of an American Civil Liberties Union public-records request, an Amazon account executive asked to introduce Adzima to an executive at a police-body-camera company who wanted to understand how he “overcame stakeholder resistance.”
“You’re AWS-famous now,” the executive wrote, with an emoji of a smiley face.
Deputies here say the system is a huge hit. Chris Lee, who has used the search in five cases of burglary and theft, said many of his colleagues have become prolific users, eager to find a simple resolution to an otherwise-difficult hunt. “You’re always like: Is it going to show us something?” he said.
For training, deputies are emailed only a printout of the office’s facial-recognition policy and a short PowerPoint presentation cautioning them to be careful with the results. One slide shows how the system responded to an uploaded mug shot of O.J. Simpson: by returning a photo of a white man with a beard. “As you can see,” the slide reads, the system “still requires human interpretation.”
The agency’s four-page policy requires staffers to use the system only in cases of a “criminal nexus” and prohibits its use in “mass surveillance” or to monitor people based on their religion, political activities or race. But it also offers several exceptions, including allowing facial searches in cases of “significant threat to life” or when deputies believe a felony suspect will be at a certain place at a specific time.
The search has helped deputies devise unconventional techniques. In one case, an inmate was talking to his girlfriend on a jailhouse phone when she said there was a warrant out for her arrest. Deputies went to the inmate’s Facebook page, found an old video with her singing and ran a facial-recognition search to get her name; she was arrested within days.
Deputies can also run black-and-white police sketches through the system looking for results; in one test case, they said, it pointed to a man they’d already flagged as their suspect. Amazon said that running sketches through Rekognition does not violate its rules but that it expects human reviewers to “pay close attention to the confidence of any matches produced this way.”
Bruington, from the county public defender’s office, said Rekognition’s low price and ease of use could tempt police agencies into experimenting with a system they may not fully understand. She also worried that the system’s dependence on mug shots meant that anyone previously brought in by police would be that much more likely to resurface in a criminal search.
“Innocent people go through the criminal justice system every day,” she said.
‘How did this work’
Facial-recognition technology had for decades been a police agency’s dream: a simple, stealthy way to identify anyone from afar, without their knowledge or consent. But only in recent years — thanks to improvements in imaging and computer power, and plunging data-storage costs — has the technology become affordable and widespread, used in tagging Facebook photos and unlocking iPhones.
Today’s systems break down people’s facial photos into long strands of code, called “feature vectors” or “faceprints,” that can be rapidly compared with other portraits across a vast database. But while “computer-vision” algorithms are adept at pattern recognition, they match pixels, not clues, and can miss inconsistencies that would seem staggeringly obvious to the human eye.
Still, the promise of cheap and easy identification has proved too compelling for many companies to ignore. The federal agency that assesses facial-recognition algorithms, the National Institute of Standards and Technology, recently said it had tested 127 systems from 44 companies on their “scalability to large populations” and accuracy in identifying “noncooperative subjects” photographed “in the wild.” The top-ranking algorithms, from Microsoft and the Chinese startup Yitu Technology, could match a face photo across a database of millions of images with 99% accuracy.
Amazon has previously declined to submit Rekognition for this assessment, saying the test, which studies an isolated version of the core search algorithm, wouldn’t work on its complicated cloud-based search. But an NIST official said that fact has not impeded other companies with similar searches. An Amazon official said the company had launched a “substantive” effort to “redesign critical components” of the system so it could participate.
The FBI said it ran more than 52,000 facial-recognition searches in the past fiscal year, and in 2016, researchers from the Georgetown University Law School found at least 52 state or local agencies that had at some point relied on a facial-search system built by federal contractors or surveillance firms. But Amazon has made it simple for any new police force to get started, charging a cut-rate fee based partially on the number of “faces stored.”
No federal laws govern the use of facial recognition. But a bipartisan bill introduced in the U.S. Senate in March and a proposed bill in Amazon’s home state of Washington could impose new rules that would, for instance, require companies to notify passers-by that their faces are being scanned. San Francisco leaders are expected to vote next week on a proposal, opposed by police, that would make the tech capital the first city in America to ban local agencies from using facial-recognition software.
Amazon executives say they support national facial-recognition legislation, but they have also argued that “new technology should not be banned or condemned because of its potential misuse.” FBI agents and Orlando, Florida, police say they have tested the system, and Amazon has pitched it to government agencies, including Immigration and Customs Enforcement.
Lawyers in Washington County, Oregon, said they’re just starting to see the technique show up in arrest reports, and some are preparing for the day when they may have to litigate the systems’ admissibility in court. Marc Brown, a chief deputy defender working with Oregon’s Office of Public Defense Services, said he worried the system’s hidden decision-making could improperly tilt the balance of power: Human eyewitnesses can be questioned in court, but not this “magic black box,” and “we as defense attorneys cannot question, you know, how did this process work.”
The system’s results, Brown added, could pose a huge confirmation-bias problem by steering how deputies react. “You’ve already been told that this is the one, so when you investigate, that’s going to be in your mind,” he said. “The question is no longer who committed the crime, but where’s the evidence to support the computer’s analysis?”
Amazon’s software is rapidly becoming more advanced. The company last month announced a Rekognition update that would, among other things, improve the accuracy of the system’s “emotion detection” feature, which automatically speculates on how someone is feeling based on how they look on camera. It includes “7 supported emotions: ‘Happy,’ ‘Sad,’ ‘Angry,’ ‘Surprised,’ ‘Disgusted,’ ‘Calm’ and ‘Confused.’ “
‘Look at the bird’
Amazon also owns Ring, the maker of a popular doorbell camera, which applied last year for a facial-recognition patent that could flag “suspicious” people at a user’s doorstep. A Ring spokeswoman said the company’s patent applications are intended to “explore the full possibilities of new technology.”
The Washington County Sheriff’s Office’s face database, meanwhile, is always growing, by roughly 19,000 jail bookings a year. When people are arrested, they’re brought to a bustling intake room where they get their picture taken by a webcam topped with a red Beanie Babies cardinal. “Look at the bird,” they’re told.
Those photos become the inmates’ identities throughout the county’s penal system, and an internal jail website and iPhone app displays the images in a large grid so deputies can quickly track their food intake, behavior and suicide risk.
Rekognition isn’t used once an inmate is in lockup, but it has nevertheless left a subtle impact behind bars. Standing in the guardhouse nerve center of Pod 3, the maximum-security wing that inmates call “the hole,” Deputy Brian van Kleef put it this way: “This is where we gather our database.”