Facial Recognition Misidentification: Wrongful Arrests, Bias, and Rules

Facial recognition misidentification happens when automated software wrongly matches a person’s face to someone else in a database, and the mistake has already put at least fourteen innocent people in U.S. handcuffs that we know about.¹ The errors trace to a mix of algorithmic weakness, demographic bias baked into training data, and police work that treats a computer’s guess as something closer to proof. No federal law governs how the technology is used, and the safeguards that do exist have not stopped the arrests from adding up.

How a Wrong Match Turns Into a Wrong Arrest

The software compares a probe image, often a grainy still from a surveillance camera, against a database of photos: mugshots, driver’s license pictures, or in some cases images scraped from the internet. It returns candidates ranked by similarity, sometimes with a probability score attached. Florida’s “FACES” system returned a 93 percent probability match in one case that turned out to be wrong.²

A high similarity score is not a positive identification, and every agency using the technology says so on paper. In practice, the algorithm’s guess often drives the investigation. In at least seven documented wrongful arrests, police took the facial recognition candidate’s photo and dropped it straight into a lineup shown to a witness.¹ The ACLU calls this “tainting” the lineup: the witness is now choosing between an innocent lookalike and random fillers, and the face that resembles the real suspect most closely is by definition the one the algorithm surfaced.

Human review is supposed to be the safeguard, and it often is not. Studies have found modern systems can outperform humans on specific face-comparison tasks, which means the officer reviewing an output may not be able to catch the machine’s mistake, and may defer to a tool they perceive as more accurate than their own eyes.³ Georgetown Law’s Center on Privacy and Technology described the whole process as “a forensic without the science,” concluding that the algorithmic and human steps compound each other’s mistakes and that facial recognition has been used as the basis for probable cause even when official policy treats it as a mere lead.⁴

Accuracy in the lab is not accuracy on the street. In cooperative conditions, like unlocking a phone, modern systems get it wrong roughly one time in a thousand. In the uncontrolled conditions of a criminal case, with poor lighting and surveillance-quality footage, accuracy drops substantially and is harder to measure, and demographic biases grow larger.³

Who Gets Misidentified Most Often

The pattern of wrongful arrests is not random. A 2019 National Institute of Standards and Technology study evaluated 189 algorithms from 99 developers against 18.27 million images and found the majority produced significantly different accuracy across demographic groups.⁵

The false positive rate — the algorithm claiming two different people are the same person — was 10 to 100 times higher for Asian and African American faces than for Caucasian faces in many algorithms tested. False positives were also consistently higher for women than for men, and elevated for children and the elderly.⁶ For U.S.-developed algorithms tested against domestic law enforcement images, Native American and American Indian populations showed the highest false positive rates, followed by African American and Asian populations. In one-to-many identification, which is the mode police use to search a database, false positive rates were particularly elevated for African American women.⁵

Training data shapes these gaps. NIST found algorithms developed in China produced low false positive rates on East Asian faces, and algorithms built in Asian countries generally did not show the dramatic Asian-versus-Caucasian gap that U.S.-developed tools did.⁶ The most equitable algorithms in the study were also among the most accurate overall.⁵

The gap has not closed. A 2025 study sponsored by the Department of Homeland Security found certain contemporary systems detected 99.7 percent of lighter-skinned subjects but only 76 percent of darker-skinned subjects in operational scenarios.⁷ Nearly all of the earliest wrongful arrests involved Black individuals, though more recent cases have affected people of other backgrounds as well.⁸

What Wrongful Arrests Have Looked Like

The known cases share a pattern. An algorithm produces a false match. Police treat the match as something close to a positive identification, sometimes running it past a witness in a photo lineup. An innocent person is arrested, held, and eventually released after the case falls apart.

Robert Williams was arrested outside his Detroit home in January 2020 and held in jail for 30 hours after facial recognition software linked him to a 2018 watch theft at a Shinola store. Detroit police later admitted “the computer got it wrong.”⁹

Porcha Woodruff was eight months pregnant when Detroit police arrested her in February 2023 for carjacking, based on a facial recognition match drawn from an eight-year-old photo. Prosecutors dropped the charges a month later.¹⁰

Nijeer Parks was arrested and held for ten days in February 2019 after facial recognition matched him to a fake driver’s license used in a New Jersey shoplifting incident.¹¹

Angela Lipps, a 50-year-old grandmother from Elizabethton, Tennessee, had never set foot in North Dakota when U.S. Marshals arrested her in July 2025 on a Fargo warrant tied to bank fraud. West Fargo police had used Clearview AI to flag her as a “potential suspect” off a fake ID, and Fargo police accepted the finding without checking her whereabouts during the alleged crimes.¹² She spent more than five months in custody across two states before her defense team produced bank records placing her in Tennessee at the time of the fraud. Charges were dismissed December 23, 2025.¹³

Trevis Williams was arrested in April 2025 after the NYPD linked him to a February incident in a Manhattan building. Williams stands six feet two inches and weighs 230 pounds, roughly eight inches taller and 70 pounds heavier than the described suspect, and phone location data placed him about 12 miles from the scene at the time. He spent more than two days in jail before the case was dismissed. The Legal Aid Society later revealed that the NYPD’s standard Facial Identification Section initially found “no match” for the suspect, and that a separate unit in the Intelligence Division then generated the erroneous match, which Legal Aid characterized as circumventing the department’s own policies.¹⁴¹⁵

Beau Burgess was arrested in August 2025 on fraud and theft charges tied to over $4,400 in unpaid Universal Orlando hotel charges. Orlando police ran a body-camera still through Florida’s FACES database and got back a decade-old booking photo of Burgess. The actual suspect in the video had no leg tattoos; Burgess’ legs are covered in them. A hotel employee still picked Burgess from a six-person photo lineup. The state attorney’s office dropped both charges. An internal affairs investigation cleared the officers, finding they had followed policy treating facial recognition as an investigative lead.¹⁶

Others publicly identified include Michael Oliver (Detroit, July 2019); Kimberlee Williams, who spent six months in jail in 2021 after a Maryland warrant traced to a facial recognition match; Christopher Gatlin (St. Louis, August 2021); Alonzo Sawyer (Maryland transit, March 2022); Randal Quran Reid, arrested in Georgia on a Louisiana warrant based solely on a Clearview AI result and held nearly a week; Jason Killinger (Reno, September 2023); Javier Lorenzano-Nunez (Phoenix, October 2024); and Harvey Eugene Murphy Jr. (Houston).¹¹⁷

What Happens After a False Match

The consequences follow the arrest even after the charges disappear. Time in custody has ranged from roughly a day to six months in the known cases. Charges have been dropped by prosecutors, but the arrest record, the job disruption, and the fact of the jail stay remain.

Civil outcomes have been mixed. Robert Williams settled with Detroit in June 2024. The city council approved a $300,000 payment, and the police department agreed to a set of reforms: no arrests based solely on facial recognition or on lineups derived from it; lineups require independent evidence tying the suspect to the crime; the use of facial recognition must be disclosed to courts; and every case involving the technology back to 2017 must be audited. A federal court retains jurisdiction to enforce the terms for four years.⁹¹⁸

Not every plaintiff has fared as well. In August 2025, a federal judge granted summary judgment against Porcha Woodruff, finding her attorneys had not demonstrated the officer lacked probable cause at the time of the arrest. An appeal is planned.¹⁹ Parks v. McCormac remains in litigation.¹¹ The ACLU filed a federal lawsuit in June 2026 on behalf of Robert Dillon, a Fort Myers resident arrested after Florida’s FACES system returned a 93 percent match tying him to a child-luring incident 300 miles away; the suit alleges officers relied on the algorithm instead of investigating exculpatory evidence like license plate reader data.²

The Rules Police Are Supposed to Follow

Nearly every law enforcement agency using facial recognition officially describes results as “investigative leads” rather than positive identifications. The NYPD’s 2026 policy requires a multi-stage human review: a trained investigator manually compares facial features, a peer reviews the finding, and a supervisor approves. The department does not run real-time facial recognition, and probe images are only compared against arrest and parole photo repositories.²⁰ The Trevis Williams case then showed a separate NYPD unit generating a match after the standard unit found none, raising questions about whether these policies are consistently followed.¹⁵

Federal agencies have been looser still. A 2023 GAO review found only two of seven federal agencies required staff to complete training before using facial recognition services, and four of the seven had no specific policies to protect civil rights and civil liberties.²¹ The U.S. Commission on Civil Rights reported in September 2024 that federal guidelines and oversight had “lagged behind” real-world use, and that no federal law expressly regulates the government’s use of the technology.²²

Where the Law Stands

Regulation has come from states and cities. As of late 2024, fifteen states had enacted laws restricting police use of facial recognition in some way. Maine, Massachusetts, Montana, and Utah require a warrant or probable cause before a search. Alabama, Colorado, Maine, Maryland, Montana, Virginia, and Washington prohibit arrests based solely on a facial recognition result. Colorado and Virginia require testing and accuracy standards. Colorado, Maryland, Montana, New Jersey, and Washington require defendants to be notified when the technology was used in their case.²³

A handful of cities have banned government use outright, including San Francisco, Oakland, and Somerville, Massachusetts.²⁴ Officers in jurisdictions without bans have reportedly sidestepped neighboring cities’ restrictions by asking uncovered agencies to run the query for them.²⁵

Federal bills have been introduced but none has become law. Congress is currently considering the ICE Out of Our Faces Act, which would bar ICE and CBP from using the technology, along with broader proposals including H.R. 3782 (federal identity verification) and H.R. 3060 (federally assisted housing).²⁶⁷

For readers comparing U.S. rules to European ones: the EU AI Act, whose prohibitions took effect in February 2025, bans real-time facial recognition in publicly accessible spaces for law enforcement, subject to narrow public-safety exceptions with judicial or administrative approval, and absolutely prohibits building facial recognition databases through untargeted scraping of the internet or CCTV.²⁷ The U.S. has no equivalent.

Misidentification Outside Policing

False matches also occur in retail settings, where consequences are different but real. The Federal Trade Commission acted against Rite Aid in December 2023 after finding that the chain deployed facial recognition in hundreds of stores between 2012 and 2020 without testing for accuracy or bias. The FTC’s complaint alleged the system generated thousands of false positive matches that “falsely and disproportionately” identified people of color and women as shoplifters, with employees responding by following customers, searching them, or calling police. The technology produced more false positives in stores in predominantly Black and Asian communities. The FTC banned Rite Aid from using the technology for five years and required deletion of all collected images and derived algorithms.²⁸

The pattern holds across settings. A false match produces a downstream human decision — an arrest, a search, an ejection from a store — and the person on the receiving end has to prove they are not the person the algorithm named. The list of people who have had to do that continues to grow.

1
ACLU. More Than a Dozen Wrongful Arrests Due to Police Reliance on Facial Recognition Technology
2
The Guardian. Florida Lawsuit AI Facial Recognition
3
Federation of American Scientists. Face Recognition Bias
4
Georgetown Law Center on Privacy and Technology. A Forensic Without the Science: Face Recognition in U.S. Criminal Investigations
5
NIST. NIST Study Evaluates Effects of Race, Age, Sex on Face Recognition Software
6
NIST. NISTIR 8280 – Face Recognition Vendor Test Part 3: Demographic Effects
7
Congressional Research Service. Face Recognition Technology
8
ACLU. Police Say a Simple Warning Will Prevent Face Recognition Wrongful Arrests — That’s Just Not True
9
Detroit Free Press. Man Wrongfully Arrested With Facial Recognition Tech Settles Lawsuit
10
New York Times. Detroit Facial Recognition False Arrests
11
ACLU. Parks v. McCormac
12
CNN. Angela Lipps AI Facial Recognition
13
New York Times. North Dakota Facial Recognition AI Errors Bank Fraud
14
New York Times. NYPD Facial Recognition Dismissed Case
15
Legal Aid NYC. NYPD POST Act Violations
16
WESH. Florida Man Wrongly Arrested Police Facial Recognition
17
ACLU. Police Say a Simple Warning Will Prevent Face Recognition Wrongful Arrests
18
ACLU. Civil Rights Advocates Achieve the Nation’s Strongest Police Department Policy on Facial Recognition Technology
19
MyNorthwest. Woman Wrongly Accused of Carjacking Loses Lawsuit Against Detroit Police Who Used Facial Tech
20
NYPD. Facial Recognition Impact and Use Policy
21
GAO. Facial Recognition Technology: Federal Law Enforcement Agencies’ Use and Views
22
U.S. Commission on Civil Rights. Civil Rights Implications of the Federal Use of Facial Recognition Technology
23
Tech Policy Press. Status of State Laws on Facial Recognition Surveillance
24
NYU Policing Project. General Regulations
25
Columbia Science and Technology Law Review. Facial Recognition Technology Law Enforcement
26
Office of Rep. Pramila Jayapal. Markey, Jayapal, Merkley, Wyden Introduce Bill to Ban ICE and CBP Use of Facial Recognition Technology
27
IAPP. Biometrics in the EU: Navigating the GDPR and AI Act
28
FTC. Rite Aid Banned From Using AI Facial Recognition