A ranked list of candidates was enough to arrest the wrong man

The search that named Robert Williams

In October 2018, someone stole several watches from a Shinola store in downtown Detroit. The store's surveillance camera caught the theft, and the still frame investigators pulled from the footage was too degraded for a confident human description of the suspect's face.¹

Detroit police ran that image through facial-recognition software supplied by DataWorks Plus, searching it against a database of driver's-license and mugshot photographs. The system returned a set of candidates ranked by how closely each one resembled the probe image. Robert Williams's expired driver's-license photo was among them.¹

That is what an identification search is built to produce: not a confirmed identity, a shortlist. The federal agency that tests these systems, the National Institute of Standards and Technology, defines the output plainly. An identification algorithm returns "a fixed number of the most similar candidates, or only those that are above a preset threshold," and each candidate is "an index and a similarity score."² A rank, not a verdict.

Thirty hours over a match nobody verified

Detroit investigators built a photo lineup around Williams's license photo and showed it to a loss-prevention contractor for Shinola. The contractor had not been at the store during the theft. He was matching a face from a printed photo array to a crime he had never witnessed.¹

On January 9, 2020, officers went to Williams's home in Farmington Hills and arrested him, in front of his wife and his two young daughters.³ He spent the next 30 hours in custody, much of it in what the ACLU's complaint called a "cramped and filthy cell," before officers showed him the surveillance still and the photo lineup during questioning.³ Williams looked at the still, then at his own photo beside it, and told them: "This is not me. I hope y'all don't think all black people look alike." The detective in the room answered: "The computer says it's you."⁴

The lawsuit Williams filed fourteen months later alleged a second failure sitting on top of the first: that the detective who sought his arrest warrant left out how unreliable the facial-recognition match was when asking a magistrate to sign it. What exactly the warrant application disclosed has not been made public.¹

Charges were still open when the ACLU filed an administrative complaint on Williams's behalf on June 24, 2020, demanding the case be dropped, the department apologize, and Detroit stop using facial recognition altogether.³ Detroit didn't drop the technology. On April 14, 2021, Williams sued the city in federal court, alleging violations of the Fourth Amendment and Michigan's civil-rights law. The suit also argued the department had no written policy governing facial recognition and had never trained officers on its documented, unequal error rates.¹ MIT Technology Review called it the first federal lawsuit to treat a face-recognition misidentification as a civil-rights case rather than a technical glitch.⁵

The case settled on June 28, 2024. Detroit paid Williams $300,000 and agreed to what the ACLU calls the strongest police facial-recognition policy in the country. Officers can no longer arrest someone, or build a photo lineup, on an unconfirmed facial-recognition lead alone. A search now requires independent corroborating evidence and three separate approvals before an arrest can follow, and a federal court will monitor compliance for four years.⁶ "They thought this was some type of magic they were using," Williams said after the settlement. "And it didn't make sense at all."⁷

The odds NIST had already measured

A year before Williams's arrest, in December 2019, NIST published results from testing 189 face-recognition algorithms submitted by 99 developers against roughly 18 million photographs drawn from four federal image sets.²

For one-to-one verification, matching a live photo to a single claimed identity the way a phone unlocks, the study found false-positive rates ran 10 to 100 times higher for Black and East Asian faces than for white faces across many of the algorithms tested.⁸

Detroit's search on Williams's photo was a different comparison: one face checked against an entire mugshot database, the one-to-many identification search NIST tested separately. There the pattern held with a twist. Across domestic mugshot searches, the highest false-positive rates showed up in American Indian faces, with African American and Asian faces elevated too. Which group came out worst depended on the algorithm, and on whether the face was male or female.²

NIST went further, and closer to what happened to Williams. A search returns similarity scores next to each ranked candidate, the format an investigator actually sees. The report found scores ran systematically higher for Black women than for other groups, an order-of-magnitude increase in several algorithms, though not all.² The report named the risk: a human reviewer working down that list would see a stronger-looking match sitting near the top, in exactly the population where the algorithm was most likely to be wrong.²

NIST's report does not name DataWorks Plus, or identify which algorithm Detroit's search used, among the systems it tested, so it cannot say what error rate rode into Williams's case that day. What it established is the pattern, not the individual number: across nearly 200 algorithms, this kind of demographic gap was common enough to call predictable.² Williams's arrest was not proof that his specific search carried those specific odds. It was an outcome NIST's numbers had already made more likely than the technology's reputation for accuracy implied.

None of that explains the arrest by itself. A degraded probe image made any match less certain to begin with. A demographic gap in the algorithm's errors made a false positive more likely for a face like Williams's. And an investigator who treated "the computer says it's you" as confirmation, rather than as one candidate among several carrying a similarity score, is what turned a statistical tail risk into a warrant. Remove any one of the three, and Williams likely never sees a jail cell.

Three more names, the same lead

Williams is described as the first publicly documented wrongful arrest caused by facial recognition in the United States. A case settled or dropped quietly would leave no public record at all.³ Kashmir Hill, the reporter who first wrote about Williams's arrest, spent the following years tracing how far the same kind of search had spread through American policing.⁹

Nijeer Parks was arrested in New Jersey in 2019 after a facial-recognition search tied him to a shoplifting and a fake ID he says he never used; he was 30 miles from the store when it happened and had never held a driver's license the state could have matched against. He spent 10 days in jail and about $5,000 defending himself before the case was dismissed for lack of evidence.¹⁰

Porcha Woodruff was eight months pregnant when six Detroit officers arrested her, in front of her children, on a carjacking charge again built on a facial-recognition lead. She was held 11 hours, had contractions in custody, and the charges were dropped a month later.¹¹

Randall Reid was arrested in Georgia in November 2022 on warrants from a Louisiana parish he says he had never visited. He spent 6 days in jail before the warrants were withdrawn and settled a federal civil-rights suit for $200,000 in 2025.¹²

Each case runs Williams's mechanism again. A probe image the software could not match with confidence, a demographic error rate testing had already shown was uneven, and an officer acted on the ranked list's top result without independent confirmation. The ACLU has counted at least seven confirmed cases; six of the seven were Black.¹³

Detroit is the only department bound by a settlement that requires independent evidence before an arrest and bars a lineup built from an unconfirmed facial-recognition lead. Parks was arrested in New Jersey. Reid's warrant originated in Louisiana. The search that produced Williams's arrest is still standard practice everywhere that settlement does not reach.