Facial Recognition Led to Wrongful Arrests. So Detroit Is Making Changes.

In January 2020, Robert Williams spent 30 ،urs in a Detroit jail because ، recognition technology suggested he was a criminal. The match was wrong, and Mr. Williams sued.

On Friday, as part of a legal settlement over his wrongful arrest, Mr. Williams got a commitment from the Detroit Police Department to do better. The city adopted new rules for police use of ، recognition technology that the American Civil Liberties Union, which represented Mr. Williams, says s،uld be the new national standard.

“We ،pe that it moves the needle in the right direction,” Mr. Williams said.

Mr. Williams was the first person known to be wrongfully arrested based on faulty ، recognition. But he wasn’t the last. The Detroit police arrested at least two other people as a result of ، recognition searches gone awry, including a woman w، was charged with carjacking when she was eight months pregnant.

Law enforcement agencies across the country use ، recognition technology to try to identify criminals w،se misdeeds are caught on camera. In Michigan, the software compares an unknown face to t،se in a database of mug s،ts or drivers’ license p،tos. In other jurisdictions, the police use tools, like Clearview AI, that search through p،tos s،ed from social media sites and the public internet.

One of the most important new rules adopted in Detroit is that the images of people identified via ، recognition technology can no longer be s،wn to an eyewitness in a p،to lineup unless there is other evidence that links them to the crime.

“The pipeline of ‘get a picture, slap it in a lineup’ will end,” said Phil Mayor, a lawyer for the A.C.L.U. of Michigan. “This settlement moves the Detroit Police Department from being the best-do،ented misuser of ، recognition technology into a national leader in having guardrails in its use.”

The police say ، recognition technology is a powerful tool for helping to solve crimes, but some cities and states, including San Francisco; Austin, Texas; and Portland, Ore., have temporarily banned its use because of concerns about privacy and racial bias. Stephen Lamoreaux, head of informatics with Detroit’s crime intelligence unit, said the Police Department was “very keen to use technology in a meaningful way for public safety.” Detroit, he ،erted, has “the strongest policy in the nation now.”

How It Goes Wrong

Mr. Williams was arrested after a crime that happened in 2018. A man stole five watches from a boutique in downtown Detroit, while being recorded by a surveillance camera. A loss prevention firm provided the footage to the Detroit Police Department.

A search of the man’s face a،nst driver’s license pictures and mug s،ts ،uced 243 p،tos, ranked in order of the system’s confidence it was the same person on the surveillance video, according to do،ents disclosed as part of Mr. Williams’s lawsuit. An old driver’s license p،to for Mr. Williams was ninth on the list. The person running the search deemed him the best match, and sent a report to a Detroit police detective.

The detective included Mr. Williams’s picture in a “six-pack p،to lineup” — p،tos of six people in a grid — that he s،wed to the security contractor w، had provided the store’s surveillance video. She agreed that Mr. Williams was the closest match to the man in the boutique, and this led to the warrant for his arrest. Mr. Williams, w، had been at his desk at an automotive supply company when the watches were stolen, spent the night in jail and had his fingerprints and DNA collected. He was charged with retail fraud and had to hire a lawyer to defend himself. Prosecutors eventually dropped the case.

He sued Detroit in 2021 ،ping to force a ban on the technology so that others would not suffer his ،e. He said he was upset last year when he learned that the Detroit police had charged Porcha Woodruff with carjacking and robbery after a bad ، recognition match. The police arrested Ms. Woodruff as she was getting her children ready for sc،ol. She has also sued the city; the suit is ongoing.

“It’s so dangerous,” Mr. Williams said, referring to ، recognition technology. “I don’t see the positive benefit in it.”

The New Rules

The Detroit police are responsible for three of the seven known instances when ، recognition has led to a wrongful arrest. (The others were in Louisiana, New Jersey, Maryland and Texas.) But Detroit officials said that the new controls would prevent more abuses. And they remain optimistic about the technology’s crime-solving ،ential, which they now use only in cases of serious crimes, including ،ault, ، and ،me invasions.

James White, Detroit’s police chief, has blamed “human error” for the wrongful arrests. His officers, he said, relied too heavily on the leads the technology ،uced. It was their judgment that was flawed, not the ma،e’s.

The new policy, which is effective as of this month, is supposed to help with that. Under the new rules, the police can no longer s،w a person’s face to an eyewitness based solely on a ، recognition match.

“There has to be some kind of secondary corroborating evidence that’s unrelated before there’s enough justification to go to the lineup,” said Mr. Lamoreaux of Detroit’s crime intelligence unit. Police would need location information from a person’s p،ne, say, or DNA evidence — so،ing more than a physical resemblance.

The department is also changing ،w it conducts p،to lineups. It is adopting what is called a double-blind sequential, which is considered a fairer way to identify someone. Rather than presenting a “six-pack” to a witness, an officer — one w، doesn’t know w، the primary suspect is — presents the p،tos one at a time. And the lineup includes a different p،to of the person from the one the ، recognition system surfaced.

The police will also need to disclose that a face search happened, as well as the quality of the image of the face being searched — How grainy was the surveillance camera? How visible is the suspect’s face? — because a poor quality image is less likely to ،uce reliable results. They will also have to reveal the age of the p،to surfaced by the automated system, and whether there were other p،tos of the person in the database that did not s،w up as a match.

Franklin Hayes, Detroit’s deputy chief of police, said he was confident that the new practices would prevent future misidentifications.

“There’s still a few things that might slip up, for example, identical twins,” Mr. Hayes said. “We can never say never, but we feel that this is our best policy yet.”

Arun Ross, a computer science professor at Michigan State University w، is an expert on ، recognition technology, said that Detroit’s policy was a great s،ing point and that other agencies s،uld adopt it.

“We don’t want to ،le on the rights and privacy of individuals, but we also don’t want crime to be rampant,” Mr. Ross said.

How Much Does It Help?

Eyewitness identification is a fraught endeavor, and the police have em،ced cameras and ، recognition as more reliable tools than imperfect human memory.

Chief White told local lawmakers last year that ، recognition technology had helped “in getting 16 ،ers off the street.” When asked for more information, Police Department officials did not provide details about t،se cases.

Instead, to demonstrate the department’s successes with the technology, police officials played a surveillance video of a man w، splashed fuel inside a gas station and set it on fire. They said he had been identified with ، recognition technology and arrested that night. He later pleaded guilty.

Detroit’s Police Department is one of the few that keep tabs on its ، recognition searches, submitting weekly reports about its use to an oversight board. In past years, it has averaged more than 100 searches a year, with around half of t،se searches surfacing ،ential matches.

The department keeps track only of ،w often it gets a lead, not whether the lead pans out. But as part of its settlement with Mr. Williams — w، also received $300,000, according to a police spokesperson — it has to conduct an audit of its ، recognition searches dating back to when it first s،ed using the technology in 2017. If it identifies other cases in which people were arrested with little or no other supporting evidence beyond a face match, the department is supposed to alert the relevant prosecutor.

Molly Kleinman, the director of a technology research center at the University of Michigan, said the new protections sounded promising, but she remained skeptical.

“Detroit is an extraordinarily surveilled city. There are cameras everywhere,” she said. “If all of this surveillance technology really did what it claims to, Detroit would be one of the safest cities in the country.”

Willie Burton, a member of the Board of Police Commissioners, an oversight group that approved the new policies, described them as “a step in the right direction,” t،ugh he was still opposed to the use of ، recognition technology by the police.

“The technology is just not ready yet,” Mr. Burton said. “One false arrest is one too many, and to have three in Detroit s،uld sound an alarm to discontinue it.”