The alarming face of facial recognition

One September day a few years back, two undercover detectives were cruising around Jacksonville, Fla., looking to build cases against local drug dealers. One of them was jotting down a few notes describing a drug buy that had just happened when all of a sudden a man popped up near the passenger-side window and asked the cops if they were good—as in, were they looking to buy drugs. The narcotics detectives said it all happened so quickly they didn’t even have time to turn on their secret video recording gear. Instead, one of them pretended to make a call with a cell phone, which served as an excuse to snap a few pictures of the guy who purportedly brought them what they’d just asked for: fifty hard.

A few photos—shot at an angle—and a name: Midnight. That’s what the detectives had to work with as they sought to identify the guy who’d sold them the $50 worth of crack cocaine. Over a week later, they passed their photos to a Jacksonville Sheriff’s Office crime analyst who tried a few different searches of police databases. She tried searching for Midnight. No luck. Then she uploaded the photos into a facial recognition system. Finally, a hit. Several, in fact. One of them was of Willie Allen Lynch, a 46-year-old man who was listed as “vagrant” on his arrest report. (He would go on to report zero income in the lead up to his trial.) Lynch was also someone who’d spent years in prison for drug sales. The detectives felt like they had their man.

Lynch was convicted in a day-long trial in May 2016. The undercover detectives identified him in court as the man who’d sold them the drugs, not mentioning the specifics of the facial recognition system that had brought him to their attention in the first place. But the conviction didn’t stop Lynch from questioning how the Jacksonville Sheriff’s Office used facial recognition technology. He’s appealing to the state Supreme Court, calling out the lower courts for denying him a look at the other photos the system turned up, the other potential Midnights out there. Court watchers like the American Civil Liberties Union (ACLU) are keeping close tabs on the case and the challenge it poses to widespread and largely unregulated police use this powerful technology. And the issue isn’t just playing out in the legal system; Lynch’s case is happening just as cities across the country have banned or are considering banning their police departments from using facial recognition technology. The US House of Representatives has also begun hearings on the issue.

In 2016, Georgetown University’s Center on Privacy & Technology reported that facial recognition networks have data on 117 million US adults. State and local police agencies across the country use the technology, which often includes access to driver’s license photos—images of tens of millions of people, most of whom likely have no criminal or arrest record. Half of all adults in the country are searched this way, according to the Georgetown report. During a recent House committee hearing, Elijah Cummings, a senior lawmaker, said the “technology is evolving extremely rapidly without any [real] safeguard. Whether we’re talking about commercial use or government use, there are real concerns about the risks this technology poses for our civil rights, and liberties, and our right to privacy.”

Ban by the bay. San Francisco became the first city to ban city agency use of facial recognition technology on May 14. It’s a trend that could be spreading. Later this week, Somerville, Mass., City Councilor Ben Ewen-Campen hopes to help pass the first East Coast ban, preventing his city from using a technology he thinks the public has yet to fully understand. “As soon as you start talking about facial recognition, it sounds almost as if you’re being hyperbolic or dystopian. I really don’t think we are,” he said. “Folks have gotten used to the idea that there are surveillance cameras. I don’t think folks have gotten used to the idea that all of that footage could be analyzed in real time to track every single person in public matched against a database of basically every person and basically give that power to the government with, at this point, no regulation, no oversight, no accountability, no transparency.”

Studies have shown that some systems developed by the country’s leading tech companies are less accurate when it comes to characterizing images of people of color or women than when the technology is applied to white men. MIT researcher Joy Buolamwini found that programs produced by major companies for judging the gender of faces were pretty accurate for white men, but their precision dropped way off for darker-skinned women. The ACLU tested photos of members of Congress on Amazon’s Rekognition software, a facial recognition system the company is marketing to police departments. The system wrongly paired 28 members to photos in criminal mug shot databases. The false positives disproportionally involved non-white members.

The technology has improved. Buolamwini did a follow-up study and found the companies’ software had gotten better at judging gender. Ewen-Campen, however, said his concerns about facial recognition go beyond whether companies can make a perfectly accurate technology: “I see the real danger being just kind of changing the fundamental power structure between the government and public with no guardrails like this. I don’t think that this is the society that most folks want to live in.” If the city of Somerville didn’t stop the technology from proliferating, every resident, every visitor would essentially be walking around with a barcode ID on their chest, he said.

People visiting an addiction counselor, a mosque, an underground political meeting, or an abortion clinic might think twice about doing so, Ewen-Campen said. “I think we are naïve to think that people would feel comfortable doing those things if they thought a completely unaccountable government agency had access to their every movement, every time they stepped out in public,” he said.

Facial recognition and police practices. The way police departments use facial recognition now is often shrouded in secrecy. According to the Georgetown researchers,  Florida police departments run an average of 8,000 monthly searches on the system that the Jacksonville Sheriff’s Office used to identify Lynch, a system run by the sheriff’s office in another Florida county, Pinellas.  As of 2016, the public defender’s office there had “never received any face recognition information” as part of a so-called Brady disclosure, the process in which prosecutors are required to turn over evidence that could potentially aid a defendant.

That tendency to keep facial recognition information out of court appears to be the case at the federal law enforcement level, as well. At the recent House committee hearing, Missouri Rep. Wm. Lacy Clay asked Kimberly J. Del Greco, a high-ranking FBI official, if criminal defendants were told that facial recognition analyses had led investigators to them. “As part of a criminal investigation, I don’t believe that’s part of the process,” she said. Just as the prosecutors in Lynch’s case wouldn’t turn over photos of other potential suspects to Lynch, the FBI, Del Greco said, doesn’t send the pictures of other people that turn up in a facial recognition search to defendants.

In San Francisco, a community activist who’s spoken out against the facial recognition ban there, Joel Engardio, thinks that while there should be policies that prevent law enforcement from abusing the technology, facial recognition shouldn’t be banned outright. Burglaries and car break-ins are at epidemic proportions in the city, he said. Police could go “through hundreds or thousands of mug shots to see if a match comes up faster. Doesn’t mean that if the match happens that that’s the person; there still has to be old fashioned police work that happens,” he said.

Engardio is vice president of the community group Stop Crime SF. The group runs a program called Court Watch that sends volunteers to court to make sure that judges, prosecutors, and other officials know that there is public interest in property crime cases. Engardio sees a parallel between Court Watch and how police departments could use facial recognition technology. “When we determine which cases to follow on Court Watch, we don’t even take cases where it’s a first-, second-, or even third-time offender. What we’re looking at is the person who’s broken into a car 20 times and has gotten off every time.”

The various cogs in the justice system—prosecutors and public defenders, to name two—have a say in how technology gets used, Engardio said. “Just because a technology can very efficiently identify every person who’s caught committing a crime doesn’t mean they’re all going to get prosecuted, and they’re all going to end up in jail. But it’s an efficient way to find the suspects,” he said. “The criminal justice system can figure out from there.”

But Lynch’s case shows that law enforcement is not showing much restraint when it comes to facial recognition. The narcotics detectives that allegedly bought drugs from Lynch had just used their audio/visual recording gear to capture a deal minutes before they encountered Midnight. They had been trawling a “high crime” area of Jacksonville, a part of town rife with prostitution, drugs, and violent crime, they said.

During a deposition, one of the detectives, whose names are redacted in court records, said coming across a drug dealer was “just an everyday thing.” He explained: “It’s just right around – make contact with people. We happened to make contact with [Lynch] around [the] 23 and Laura area. We drove by. He hollered at us, said, “Hey, y’all good?”

It’s hard to say whether Midnight and Lynch are the same person. As the appellate court that ruled against Lynch noted, Lynch’s jury saw the detective’s photos of Midnight, confirmed photos of Lynch, and Lynch himself. Lynch wanted to have the crime analyst, the woman who’d run the facial recognition search that identified him, testify. The program gave stars to results to indicate the closeness of the match; there had only been one star next to Lynch’s photo. “Her testimony is very, very vital to my defense. It will exonerate me,” Lynch told the judge. The judge said that the analyst’s testimony would only help the case against Lynch. Lynch’s public defender agreed. (Lynch frequently clashed with his lawyers and repeatedly asked to represent himself.) A key point Lynch’s appeal to the Florida Supreme Court tries to make is that Lynch should have had access to the other matches the analyst turned up using the facial recognition search. Lynch’s “sole defense at trial was misidentification,” Lynch’s new lawyer wrote. “The photos of the other potential matches were exculpatory, and Petitioner should have been able to present those in his defense at trial.”

Will facial recognition fill prisons with minor offenders? Even if the facial recognition system was right and Lynch is Midnight, some observers question whether facial recognition should be used to make it even easier for authorities to prosecute peopled who’ve committed relatively low-level crimes. Human Rights Watch researcher Sarah St. Vincent told Slate that if governments were going to use a powerful technology like facial recognition, “they should only be considering that in cases that are similarly serious.”

Jacksonville detectives spent maybe five minutes interacting with Lynch, probably less. They waited maybe a half-an-hour to get a facial recognition hit on him. A pretty good return on investigative time investment. At the same time, the result was yet another indigent addict, thrown in prison. Even if he was a drug dealer, Lynch was certainly no kingpin; he was just one of the numerous people the Jacksonville detectives knew they would inevitably “contact” if they drove around north of downtown long enough.

During Lynch’s sentencing hearing in June 2016, Assistant State Attorney Brett Mereness told the court that not that long before Lynch was arrested, he had walked out of prison, having served a 16-year stint for selling cocaine. This time, Mereness wanted a 30-year sentence. For Lynch, that would mean potentially leaving prison in his late 70s.

But Judge Mark Borello had something else in mind: “The vast majority of sale cocaine cases that have come through the system, especially in the [Repeat Offender Court] divisions, for many, many, many, many years have been addicts,” he told the court. “And that’s always bothered me.” Borello gave Lynch an eight-year sentence and ordered him to get addiction treatment: “Everything in your record relates to your use of drugs. I have no doubt in my mind, after having dealt with you for several months, that had you not had this issue you would have had a different life.”

When the Jacksonville detectives initially came across Midnight, it was on a dead-end street lined with abandoned houses. He was in front of what one detective called a “no-name” apartment. Before running Midnight’s photo through a facial recognition program, detectives didn’t have Lynch’s name either. Lynch’s case raises the question of whether police can responsibly use a powerful technology that can get it wrong. Activists around the country think they already know the answer: no.