Do Masks Stop Facial Recognition? What the Testing Found

September 29, 2026

Last reviewed October 20, 2026

By Dark Flock

Updated October 20, 2026

We sell balaclavas, so this is the right place to be plain about them. A face covering makes a face recognition system miss more often. It does not make you anonymous. In some states the mask is the legal problem, and the camera is the smaller one.

How we sourced this: primary documents first (NIST reports, court opinions, statutes, settlements, and government policies), then reporting. Vendor claims are labeled as the vendor's, and anything only reported is marked as such.

How does facial recognition work?

Four steps. Software finds a face in an image (detection), lines it up (alignment), converts it into a string of numbers (a template, or embedding), and compares that template with others to produce a similarity score. If the score clears a threshold someone chose, the system calls it a match. The Seventh Circuit, describing Clearview AI this July, called the output "facial vectors" reflecting the geometry of a face.

Two kinds of error follow. A false non-match misses the right person. A false match pairs two different people. Moving the threshold trades one for the other.

What is the difference between 1:1 and 1:N face recognition?

One-to-one (1:1) verification asks whether you are the person on this ID. A TSA podium camera does this. One-to-many (1:N) identification asks who you are, searching a gallery of mugshots, license photos, or scraped images. Police investigations and Clearview are 1:N. In a 1:N search a false match isn't a locked door. It's an innocent person on a list of suspects.

Do masks stop facial recognition? What NIST found

The National Institute of Standards and Technology (NIST) measured this directly, which is more than most arguments on the subject can say.

NISTIR 8311: pre-pandemic algorithms (July 2020)

NIST took 89 algorithms built before COVID-19 and ran 6.2 million photos of 1 million people through them, with masks drawn on by software. The unmasked reference photos came from immigration applications. The masked probe photos came from border crossings. This was 1:1 verification.

Without masks, the most accurate algorithms failed to match about 0.3 percent of people. With the highest-coverage mask, about 5 percent. Many algorithms that did well on bare faces failed on 20 to 50 percent of masked images. The masks covered about 70 percent of the face. Wide masks caused about twice the misses of rounder N95-style ones, and most algorithms did worse with black masks than light blue. False matches went down slightly.

NISTIR 8331: the algorithms adapted (November 2020, updated January 2022)

Vendors responded quickly. The final January 2022 update covered 319 algorithms, 266 of them submitted after mid-March 2020. With the best algorithms, the miss rate rose from 0.2 to 0.4 percent unmasked to 1 to 3 percent masked. With a typical medium-coverage wide mask, the best did about 1 to 2 percent. NIST put masked performance at roughly the state of the art for bare faces in mid-2018. Many algorithms still missed 10 to 40 percent.

One finding cuts the other way. When both photos were masked, most algorithms produced false matches 10 to 100 times more often. Masked strangers start to look alike to the software.

What NIST's mask tests did not measure

NIST listed the limits itself. The masks were digital and a single uniform color. Subjects were cooperative and mostly looking at the camera. Only 1:1 algorithms were tested. Real masks, eyewear, and patterned fabric weren't tested, and NIST didn't break results out by demographic group. Human reviewers, who in real systems fix detection errors and adjudicate candidates, weren't part of it. No balaclavas. No street cameras.

NIST also noted a suggestion that a distinctive mask worn every day could act as a "soft biometric," making the wearer easier to pick out. In February 2024 it stopped running the mask benchmark. The January 2022 report is its last published mask study, and the results table was last updated in March 2024.

Has masked face recognition improved since 2022?

Probably, but nobody has published a fresh independent measurement. Masked-face and periocular (eye-region) recognition became their own research field, with surveys in 2023 and 2024. Lab results on curated datasets are not field results. Clearview's homepage claims "99+% accuracy for all demographics." That's a company claim.

Is facial recognition less accurate for some people? NISTIR 8280

NIST's December 2019 demographic study ran 18.27 million images of 8.49 million people through 189 mostly commercial algorithms. Its main finding: false positive rates often varied by factors of 10 to beyond 100 across demographic groups. False negatives varied less, usually by factors below three.

On high-quality application photos, false positives were highest for West and East African and East Asian people and lowest for Eastern Europeans, though several algorithms developed in China reversed the East Asian result. On U.S. mugshots, false positives were highest for American Indians, with elevated rates for African American and Asian people. Women had higher false positive rates than men. So did the oldest and youngest subjects.

NIST also found some highly accurate 1:N algorithms with no detectable false positive differences, and told system owners to "know their algorithm." A city's results depend on which one it bought.

What a mask doesn't hide

• How you walk. In 2018 the AP reported that Chinese police were testing Watrix gait recognition. The company said it could identify people from up to 50 meters away with faces covered. Company claims, years old.
• Your clothes and build. MIT Technology Review reported in May 2025 that Veritone's Track, used by about 400 customers including police departments, tracks people across video by body size, gender, hair, clothing, shoes, and accessories. It doesn't use faces at all. ACLU attorney Nathan Wessler noted that someone who wears the same boots, coat, and backpack every winter day has a profile that is "the same day after day."
• Your car. Plate readers don't care about your face. See our guide to Flock cameras.
• Your phone, which reports its location to your carrier whatever is on your face.
• Photos already online. Clearview's search runs on pictures scraped from the web, many taken long before you put on a mask.

Clearview AI: what courts and regulators have said

Clearview scrapes photos from public websites, including social media, and sells search. Its homepage claims "100+ Billion Images" in its law enforcement database, which it calls "the largest in the world by far." Company claim.

• ACLU v. Clearview AI (2022). The settlement under Illinois's Biometric Information Privacy Act (BIPA) permanently bars Clearview from making its faceprint database available to most private entities nationwide. It also barred sales to any Illinois entity, state and local police included, for five years, and required an opt-out for Illinois residents.
• The BIPA class action. The settlement would pay the class the equivalent of a 23 percent stake in Clearview at an IPO, sale, or similar event, or a share of revenue instead. The district court accepted a January 2024 valuation of about $225 million, which put the stake near $51.75 million. On July 13, 2026, the Seventh Circuit vacated approval because the nationwide class had no representative of its own, and remanded. The deal never included limits on how Clearview does business.
• Abroad. On October 20, 2026, Clearview was in the UK Court of Appeal fighting a £7.5 million fine and deletion order from the Information Commissioner's Office, Biometric Update reported. A Clearview attorney said 190 of its 236 clients are police, federal investigative, or military bodies.

The Seventh Circuit says that since the ACLU settlement, Clearview's customers have been only federal and state government agencies and their contractors. NBC News, citing federal spending records, reported in February that ICE holds two Clearview contracts worth about $6 million combined.

Wrongful arrests: when a match becomes a warrant

Every documented case starts the same way. Software returns a candidate. Someone treats the candidate as an answer.

• Robert Williams, Detroit. Arrested in 2020 at his home in front of his family for a watch theft he didn't commit. His June 2024 settlement bars Detroit police from arresting anyone based solely on a face recognition result, or on a photo lineup that directly follows one. It also bars lineups built only on a face recognition lead, requires training on misidentification risks, orders an audit of cases back to 2017, and keeps the court in charge for four years. The ACLU says Detroit had three known wrongful arrests tied to the technology. All three people were Black.
• Porcha Woodruff, Detroit. Eight months pregnant when she was arrested in February 2023 for a carjacking; she spent about 10 hours in jail and charges were dropped, the AP reported. A federal judge called the arrest "troubling" but dismissed her suit in August 2025. She appealed to the Sixth Circuit. We found no appellate decision as of October 20, 2026.
• Robert Dillon, Florida. Arrested in 2024 for a crime in Jacksonville Beach, a city he says he had never visited. The ACLU says police let the false match taint a photo lineup and left out of the warrant application a plate reader search that found no trace of his car nearby. He sued in June 2026.
• Angela Lipps, Tennessee. Jailed for nearly six months on a North Dakota bank fraud case after face recognition linked her to the suspect, The Guardian reported. Charges were dismissed. She sued Fargo in September 2026.

The ACLU calls the lineup step "a recipe for misidentifications": a false match looks like the suspect, and the filler photos look less like the suspect, so a witness picks the lookalike.

The newest version runs on a phone. WIRED and NBC News reported in February 2026 that immigration agents use a DHS app, Mobile Fortify, to scan faces in the street, including U.S. citizens, observers, and protesters. Illinois and Chicago say in a lawsuit that it has been used more than 100,000 times. WIRED found the app returns candidate matches, not confirmed identities, and that CBP says the data can be kept up to 15 years. DHS calls claims that the app violates the Fourth Amendment or compromises privacy "false." NBC also described a video in which an agent, phone raised, told a man to remove his hood.

Retail facial recognition: the FTC's Rite Aid order

From 2012 to 2020, according to the Federal Trade Commission, Rite Aid ran face recognition in hundreds of stores against a watchlist built partly from low-quality security and phone images. The FTC says the system generated thousands of false matches. Employees followed, searched, and ejected customers, sometimes in front of family, and false positives were more likely in stores in plurality-Black and Asian neighborhoods. In December 2023, Rite Aid agreed to a five-year ban on face recognition for surveillance, plus deletion of the images and any algorithms built from them.

Where facial recognition is banned or limited

• Cities. San Francisco, Boston, and Minneapolis have banned police use, the ACLU notes. MIT Technology Review describes San Francisco's and Oakland's bans as near-complete.
• States. Maine and Montana sharply limit police use and bar real-time use on live video. Maryland's law limits police searches to listed serious crimes and bans live identification. It bars analyzing images of people engaged in constitutionally protected activity without reasonable suspicion of a crime, and it forbids using a match as the sole basis for probable cause. A trained person must verify every result, and police may not tell a witness before a lineup that a suspect came from face recognition.
• New Orleans. The Washington Post reported in May 2025 that police had quietly used Project NOLA, a private network of live face recognition cameras, for about two years, against a city ordinance. In April 2026 the ACLU reported records suggesting officers were still sending photos to the system in November 2025, after the announced pause.

There is a gap. Laws written around "biometric" data may not reach tools like Track that follow clothing instead of faces.

Does the TSA use facial recognition, and can you opt out?

TSA uses 1:1 comparison: a camera at the podium checks your face against your ID. By TSA's account it runs on CAT-2 units at about 350 airports and is voluntary. Tell the officer before you hand over your ID, and you'll get a manual check without losing your place. TSA says photos are deleted after a match except in limited testing, travelers under 18 are not photographed, and the images are not used for surveillance or law enforcement. That is the agency describing itself.

Are face masks legal? Anti-mask laws by state

Most states don't restrict masks. The ICNL/USCNL tracker, last updated September 2025, counts 23 states and Washington, D.C. that do. The laws fall into three types: general bans with exceptions, bans tied to a specific intent, and bans or penalty enhancements tied to committing a crime.

• General bans include Alabama, Georgia, Louisiana, Minnesota, North Carolina, South Carolina, Virginia, and West Virginia. Exceptions vary. Minnesota exempts weather protection; West Virginia exempts weather and winter sport. In Virginia a violation is a felony.
• North Carolina (2024). The old health exemption became one for medical or surgical grade masks worn to prevent the spread of contagious disease. Wearers must remove the mask when a police officer asks, or briefly when a property owner asks, for identification. A new enhancement raises any crime one class if the defendant wore a mask or other clothing to conceal identity.
• New York. A 2025 state law targets masks worn to avoid identification while committing a felony or class A misdemeanor. Nassau County's 2024 Mask Transparency Act goes further: concealing your identity with a mask in public is a misdemeanor punishable by up to a year in jail and a $1,000 fine, with exceptions for the wearer's own health or safety and for religious or cultural reasons. A federal challenge was dismissed for lack of standing in 2024.
• New laws keep arriving. New Jersey added a mask provision in 2025, and Texas now requires public universities to bar identity-concealing masks during campus expressive activity in some circumstances. Modesto, California's 2026 ordinance made bandanas and gas masks at protests a misdemeanor. A federal court blocked it with a preliminary injunction in mid-2026 while the lawsuit proceeds.

These laws change often. Read your own state's statute and any local ordinance before a protest. This is not legal advice.

Why wear a face covering, then?

For the reasons people always have. Cold. Wind on a motorcycle. Sun on a long day. Keeping your face out of a stranger's video that ends up online. Our Blackout Ops Hood and Bone Mask Liner are winter and helmet-liner wear. The Ice Silk Gaiter is listed by its maker as breathable and anti-UV. The Ghost Hood Cotton is a full cotton hood. A Drizzle Dry Cap keeps rain and glare off your eyes.

None of them defeats face recognition reliably. We don't claim otherwise, and you shouldn't buy one thinking it does. In some states, wearing one at a demonstration can be a separate offense. And a distinctive hood worn every day is itself something to recognize.

What actually limits facial recognition

Fabric doesn't. Rules do. The meaningful limits so far have come from a settlement in Detroit, a statute in Maryland, an FTC order, a consent decree in Illinois, and city councils that voted no.

• Ask for your police department's face recognition policy and its search logs by public records request.
• Ask what it searches against: mugshots, license photos, or a scraped database.
• Ask whether defendants are told when face recognition produced the lead.
• Push for the Detroit and Maryland rules in writing: no arrest on a match alone, no lineups built on a match alone, no live identification, and no searches of people at protests without suspicion of a crime.

The software keeps getting better at faces. The question that matters is who may run a search, against which photos, and what has to happen before a match becomes a knock on the door.

Sources

NIST and research

• NIST, "NIST Launches Studies into Masks' Effect on Face Recognition Software" (July 27, 2020): https://www.nist.gov/news-events/news/2020/07/nist-launches-studies-masks-effect-face-recognition-software
• NISTIR 8311, FRVT Part 6A: Face recognition accuracy with masks using pre-COVID-19 algorithms (July 2020): https://nvlpubs.nist.gov/nistpubs/ir/2020/NIST.IR.8311.pdf
• NISTIR 8331, FRVT Part 6B: Face recognition accuracy with face masks using post-COVID-19 algorithms (Nov 2020): https://nvlpubs.nist.gov/nistpubs/ir/2020/NIST.IR.8331.pdf
• NIST, FRTE Part 6B report, updated Jan 20, 2022: https://pages.nist.gov/frvt/reports/facemask/frvt_facemask_report.pdf
• NIST, FRTE Face Mask Effects status page (benchmark discontinued Feb 22, 2024): https://pages.nist.gov/frvt/html/frvt_facemask.html
• NISTIR 8280, FRVT Part 3: Demographic Effects (Dec 2019): https://nvlpubs.nist.gov/nistpubs/ir/2019/NIST.IR.8280.pdf
• Alonso-Fernandez et al., "Periocular Biometrics: Databases, Algorithms and Directions" (2023): https://arxiv.org/abs/2307.14111
• Mahmoud et al., "A Comprehensive Survey of Masked Faces: Recognition, Detection, and Unmasking," Applied Sciences (2024): https://www.mdpi.com/2076-3417/14/19/8781
• EFF Street Level Surveillance, Face Recognition: https://sls.eff.org/technologies/face-recognition

Courts, settlements, and government

• In re Clearview AI, Inc. Consumer Privacy Litigation, No. 25-1673 (7th Cir. July 13, 2026): https://media.ca7.uscourts.gov/cgi-bin/OpinionsWeb/processWebInputExternal.pl?Path=Y2026%2FD07-13%2FC%3A25-1673%3AJ%3AHamilton%3Aaut%3AT%3AfnOp%3AN%3A3572523%3AS%3A0&Submit=Display
• ACLU, ACLU v. Clearview AI (settled May 11, 2022): https://www.aclu.org/cases/aclu-v-clearview-ai
• ACLU, Williams v. City of Detroit: https://www.aclu.org/cases/williams-v-city-of-detroit-face-recognition-false-arrest
• ACLU of Michigan, Williams settlement announcement (June 28, 2024): https://www.aclu.org/press-releases/civil-rights-advocates-achieve-the-nations-strongest-police-department-policy-on-facial-recognition-technology
• Woodruff v. Oliver, E.D. Mich. notice of appeal (Aug 30, 2025): https://storage.courtlistener.com/recap/gov.uscourts.mied.371448/gov.uscourts.mied.371448.80.0.pdf
• ACLU, Dillon v. City of Jacksonville Beach (filed June 10, 2026): https://www.aclu.org/cases/dillon-v-city-of-jacksonville-beach
• FTC, "Rite Aid Banned from Using AI Facial Recognition" (Dec 19, 2023): https://www.ftc.gov/news-events/news/press-releases/2023/12/rite-aid-banned-using-ai-facial-recognition-after-ftc-says-retailer-deployed-technology-without
• Maryland Code, Criminal Procedure § 2-502: https://mgaleg.maryland.gov/2026RS/Statute_Web/gcp/2-502.pdf
• Maryland Code, Criminal Procedure § 2-503: https://mgaleg.maryland.gov/2026RS/Statute_Web/gcp/2-503.pdf
• TSA, Facial Comparison Technology fact sheet: https://www.tsa.gov/news/press/factsheets/facial-comparison-technology
• TSA, Digital Identity and Facial Comparison Technology: https://www.tsa.gov/digital-id
• ICNL/USCNL, Anti-Mask Laws in the United States (updated Sept 2025): https://www.icnl.org/our-work/us-program/assembly/anti-mask-laws-in-the-united-states
• North Carolina Session Law 2024-16 (HB 237): https://ncleg.gov/EnactedLegislation/SessionLaws/HTML/2023-2024/SL2024-16.html
• NYCLU, "Your Rights to Wear a Mask in New York" (Nassau County Mask Transparency Act): https://www.nyclu.org/resources/know-your-rights/your-rights-to-wear-a-mask-in-public-in-nassau-county
• Civil Rights Litigation Clearinghouse, G.B. v. Nassau County: https://clearinghouse.net/case/45726/
• ACLU of Northern California, Pacific Media Workers Guild v. City of Modesto: https://www.aclunorcal.org/cases/pacific-media-workers-guild-et-al-v-city-of-modesto-et-al/

Company materials

• Clearview AI homepage (database size and accuracy claims): https://www.clearview.ai/

Civil liberties groups and journalism

• ACLU, "New Orleans Police Continuing to Use Live Face Recognition Despite City Law" (Apr 2026): https://www.aclu.org/news/privacy-technology/new-orleans-face-recognition
• The Washington Post, "Police secretly monitored New Orleans with facial recognition cameras" (May 19, 2025): https://www.washingtonpost.com/business/2025/05/19/live-facial-recognition-police-new-orleans/
• MIT Technology Review, "How a new type of AI is helping police skirt facial recognition bans" (May 12, 2025): https://www.technologyreview.com/2025/05/12/1116295/how-a-new-type-of-ai-is-helping-police-skirt-facial-recognition-bans/
• AP, "Chinese 'gait recognition' tech IDs people by how they walk" (Nov 2018): https://apnews.com/article/bf75dd1c26c947b7826d270a16e2658a
• AP, "Judge dismisses lawsuit against Detroit police in wrongful arrest case" (Sept 2025): https://apnews.com/article/detroit-facial-recognition-arrest-821d260e932a4582a6a912dd61fde157
• The Guardian, Angela Lipps lawsuit (Sept 18, 2026): https://www.theguardian.com/us-news/2026/sep/18/tennessee-woman-falsely-charged-bank-theft-lawsuit
• WIRED, "ICE and CBP's Face-Recognition App Can't Actually Verify Who People Are" (Feb 5, 2026): https://www.wired.com/story/cbp-ice-dhs-mobile-fortify-face-recognition-verify-identity/
• NBC News, "How ICE agents are using facial recognition technology to bring surveillance to the streets" (Feb 6, 2026): https://www.nbcnews.com/tech/security/ice-agent-facial-recognition-video-protest-movile-fortify-photo-rcna257331
• Biometric Update, "Clearview AI spars with UK ICO over facial recognition as monitoring" (Oct 20, 2026): https://www.biometricupdate.com/202610/clearview-ai-spars-with-uk-ico-over-facial-recognition-as-monitoring