Facial Recognition: Accuracy, Arrests and Rules

Face recognition software does one thing: it turns a face into a list of numbers and compares that list with other lists. Everything else, from opening a phone to naming a robbery suspect, depends on what it is compared against and what people do with the answer.

How a match happens

The software finds a face in an image, lines it up, and runs it through a trained model that outputs a template, a long string of numbers describing the face's features. Two templates get a similarity score. If the score passes a threshold, the system calls it a match.

The uses split sharply.

One-to-one verification asks "is this the same person?" Your phone compares your face to the one template you enrolled. Passport gates do the same with your passport photo. The odds of a mistake are low because there is only one comparison and the photos are usually good.

One-to-many identification asks "who is this?" A detective uploads a probe image, maybe a grainy still from a gas station camera, and the system compares it with every face in a gallery: a state's mugshots, sometimes driver's license photos. It returns a candidate list, ranked by score. A human analyst picks from that list. This is the kind of use behind every wrongful arrest on record.

	One-to-one	One-to-many
Question	Is this the person they claim to be?	Who in this gallery looks most like this face?
Comparisons	1	Thousands to millions
Typical image quality	Good, cooperative subject	Often poor: angle, blur, low light
Output	Yes or no	Ranked candidate list
Main risk	Rejecting the right person	Pointing at the wrong person

What NIST found

The National Institute of Standards and Technology runs the main independent tests of face recognition, now called the Face Recognition Technology Evaluation (FRTE). In December 2019 it published NISTIR 8280, a study of demographic effects. NIST tested 189 algorithms from 99 developers on about 18.27 million images of 8.49 million people, all drawn from State Department, Homeland Security and FBI databases.

The findings, in short:

• In one-to-one tests, many algorithms produced more false positives for Asian and African American faces than for white faces, sometimes 10 to 100 times more, depending on the algorithm.
• Among U.S.-developed algorithms, false positive rates were high for Asian, African American and Native American faces, with American Indian faces highest.
• Some algorithms developed in Asia did not show the large gap between Asian and white faces.
• In one-to-many tests against 1.6 million FBI mugshots, some algorithms produced more false positives for African American women.
• The most equitable algorithms were also among the most accurate.

That last point gets dropped a lot, by both sides. A police department can truthfully say "our vendor's algorithm ranks near the top in NIST testing." An advocate can truthfully say "NIST found large demographic gaps." Both claims can stand at once because algorithms differ enormously. So the question to ask is which algorithm, which version, and how it ranked. NIST keeps updating its demographic tables; the most recent update on its page is from March 2025.

The arithmetic of a big gallery

Wrongful arrests on record

The first publicly known case was Robert Williams. In January 2020, Detroit police arrested him at his home in front of his wife and two young children, accused of stealing watches from a store. A face recognition search had matched a blurry surveillance image to an older photo of him. He was not the man in the video. With the ACLU, he sued. The settlement, announced June 28, 2024, set rules the ACLU called the strongest in the country:

• Police may not arrest someone based solely on a face recognition result, or on a photo lineup that directly follows a face recognition search.
• A lineup based on a face recognition lead needs independent, reliable evidence linking the person to the crime.
• Officers get training on the technology's risks, including higher misidentification rates for people of color.
• An audit reviews every case since 2017 in which face recognition was used to get an arrest warrant.

The ACLU said Detroit police had relied on face recognition in at least three wrongful arrests, all of Black people. Nationally, Stateline counted at least seven people falsely arrested after face recognition matches as of early 2025. The count is a floor; defendants often are not told face recognition was used.

In each known case, the software was not the only failure. Detectives treated a candidate as a suspect and skipped the ordinary work: checking alibis, comparing height, looking at phone records. That is why the better rules focus on what must happen after the match.

Bans and rules

Several cities, beginning with San Francisco in 2019, voted to bar city departments from using face recognition at all. States have mostly chosen limits instead of bans. By the end of 2024, 15 states had some law on police face recognition, according to Stateline and the Center for Democracy and Technology: Washington, Oregon, Montana, Utah, Colorado, Minnesota, Illinois, Alabama, Virginia, Maryland, New Jersey, Massachusetts, New Hampshire, Vermont and Maine.

Type of rule	Example states	What it does
Warrant requirement	Montana (2023), Utah (2024)	A judge must approve the search, with exceptions such as emergencies
Serious crime limit	Maryland (2024), Vermont	Use allowed only for listed serious offenses
Notice to defendants	Maryland, New Jersey, Montana	Prosecutors must tell defendants face recognition was used
Not sole basis for arrest	Washington	A match alone cannot justify an arrest
Body camera ban	New Hampshire, Oregon	No face recognition on body camera footage

Idaho was not on the 15-state list. As of October 2026, this course found no Idaho statute specific to police face recognition; check the legislature's site for later bills.

Supporters of police use make a real case. Face recognition has helped identify suspects in violent crimes, child exploitation victims, and people who cannot identify themselves. The Policing Project, a group that pushes for rules on police technology, has said some of these tools can provide benefits for solving crime and protecting victims. The fight in most legislatures is over safeguards, not over whether the tool works at all.