Every couple of months, Officer Michael Zinn of the York Area Regional Police Department in Pennsylvania ran a photo of a man through the department's facial recognition software. In a year and a half of searches, the result was always the same: No hits.
The man was a suspect in the July 2016 sexual assault of a 15-year-old, but his identity remained a mystery. The photo came from the cellphone of the victim, who had exchanged text messages and pictures with the suspect before the assault.
Then last December, Zinn ran a search again, comparing the man's picture with database images of mug shots and driver's licences. This time, he got a possible match.
"It was a really good feeling," he said. "If it wasn't for facial recognition, it would still be an open case. We really didn't have a whole lot of leads to go on."
Facial recognition technology raises fears of a dystopian surveillance state, with vanishing privacy and a high potential for abuse. Such concerns led San Francisco last week to ban any use of facial recognition by the police and other city agencies.
But it is also a powerful and efficient tool that, much like DNA analysis, offers a way to bring policing into the modern age and help catch wrongdoers or solve crimes that have gone cold.
It has been used to arrest men accused of child sex abuse, including someone who had fled to Nepal and a man in Oklahoma who had been at large for two decades. It has helped nab a trio of jewel thief suspects and people who the authorities said were trying to enter the country under fake names.
The man in Pennsylvania, Robert Kusma, 33, was arrested and charged days after Zinn identified him, and the police said he admitted to the assault. He has not yet entered a plea in the case, and his lawyer did not respond to attempts to contact him.
It is difficult to say exactly how many of the nation's 18,000 police departments use facial recognition or how they deploy it. Some departments have been caught using it without the public's knowledge, or to search crowds of protesters for people with outstanding warrants.
But since the San Francisco ban, several agencies have come forward to argue that it is counterproductive to forbid any use of what they call a valuable tool that generates investigative leads.
Some departments, including the New York Police Department, have policies that say that a possible match found by facial recognition does not constitute an identification or probable cause for an arrest.
"We never make an arrest based solely on facial recognition," said Captain Chuck Cohen of the Indiana State Police and executive director of the Indiana Intelligence Fusion Center since late 2015. The state police have used it for a decade, scanning images of people against databases of pictures in state records and using algorithms to produce a list of individuals who could be a match. The Indiana State Police run the fusion center, which provides help to the more than 600 local law enforcement agencies in the state.
In one recent case, state police using facial recognition were able to identify a suspect in an attempted murder in Central Indiana, after a friend of the victim recorded an argument between the suspect and the victim that escalated into a physical attack. The victim, who was shot, did not know the assailant's identity.
"In this case, local law enforcement had a very good video image of the person but couldn't identify him," Cohen said. "This was not the only piece of evidence, but it was a lead."
In a twist, the agency has also used facial recognition to track down crime victims, this time in a case in which a man was extorting girls and women over the internet to perform sexual acts or mutilate themselves. Some of them may never have been found but for facial recognition, using video that had been collected by the man, Cohen said.
In one case in Phoenix in 2017, the police used a perpetrator's own cellphone to identify a suspect. Witnesses saw a knife-wielding assailant chase a man into the street and stab him several times. When the attacker fled, he left his cellphone behind; the police sent selfies from the phone to the state Department of Public Safety, whose facial recognition unit produced a potential match.
The suspect in the case, Roberto Santiago-Escobar, pleaded guilty to aggravated assault last year.
But there are no national guidelines for how facial recognition should be used. Several states, including Maryland and Indiana, allow the police to search large databases such as driver's license photos. But in Oregon, they search only against images collected in criminal proceedings, like mug shots. Private companies may offer to run images against photographs collected from more disparate sources, like social media.
"With regard to driver's licenses, it is an ethical dilemma because we are searching the faces of citizens who never consented to this type of search in the first place," said Chris Adzima, senior information systems analyst at the Washington County Sheriff's Office in Oregon, which has been using facial recognition technology for about two years. "When you take your photo at the DMV, it is not considered a public record."
He said that the sheriff's office had used it for cases ranging from "murder to shoplifting and in between."
The technology has been most effective in solving property crimes, like package thefts, pulling images from video doorbells and surveillance cameras, he said.
The police have used facial recognition to identify a suspect already in custody. After the killings of five employees at a newspaper in Annapolis, Maryland, last June, the police apprehended a suspect who refused to identify himself. The police used a scan through the Maryland Image Repository System of mug shots and driver's licenses, which helped reveal his name, Jarrod Ramos.
"We would have been much longer in identifying him and being able to push forward," Timothy Altomare, chief of the Anne Arundel County Police Department, said at the time.
The use of facial recognition by law enforcement to identify suspects raises particular concerns because of high inaccuracy rates and striking racial disparities in how the algorithms can correctly identify people.
And there are open questions about whether defendants are informed by prosecutors about how they were identified and given pertinent information that might be used to challenge that identification. Two new reports from the Georgetown Law Center on Privacy & Technology say facial recognition has been deployed irresponsibly by the police and conjure images of a futuristic surveillance state in Detroit and Chicago. Clare Garvie, an author of both reports, believes that a moratorium on facial recognition is necessary, given the lack of regulation around the technology.
"There is a fundamental absence of transparency around when and how police use face recognition technology," Garvie said. "The risks of misidentification are substantial."
In a case in New York detailed in one of the reports, a police detective noted that a suspect resembled actor Woody Harrelson, so he ran a picture of Harrelson through the facial recognition software. "This celebrity 'match' was sent back to the investigating officers, and someone who was not Woody Harrelson was eventually arrested for petit larceny," the report said.
The New York Police Department declined to say whether the man had been convicted but defended its use of the technology, saying that all leads were verified through investigation and that facial recognition had led to arrests in homicides, rapes and robberies and had even helped identify a woman with Alzheimer's.
"The NYPD has been deliberate and responsible in its use of facial recognition technology," the department said in a statement. "We compare images from crime scenes to arrest photos in law enforcement records. We do not engage in mass or random collection of facial records from NYPD camera systems, the internet or social media."
One of the Georgetown reports, relying on documentation from the department, cited two cases in which potential suspects were identified by facial recognition software — one person arrested based solely on a possible match, and a second individual identified by a facial recognition comparison and then a text message to the crime victim. In neither case, the report states, was a lineup or in-person identification conducted.
According to internal police department documents obtained by Garvie, from 2010 through 2016, more than 2,800 arrests were made as a result of facial recognition comparisons. In 2018 alone, more than 8,000 cases involved the use of facial recognition as an investigative tool.