What problem does it solve?
Police forces and retailers each keep a short list of people they want to find or keep out: wanted offenders, people in breach of court conditions, people who have seriously harmed staff. Spotting them in a crowd or a queue by eye does not work, and officers or guards cannot memorize hundreds of faces. Retrospective matching has a similar problem: investigators hold images from cameras and must search a large database by hand.
Facial recognition automates the comparison, and that creates a different problem. A wrong match leads to a person being stopped, followed or searched. The regulator reports on this page look at the safeguards around the software: the lawful basis for the watchlist, the alert threshold, human review of every alert, deletion of the faces that do not match, and monitoring for bias. Of the deployments on this page, two passed a regulator's audit or inquiry (neither regulator called it a general approval), and one ended in a five year ban under a settlement with the US Federal Trade Commission.
How does it work?
- Build a watchlist. The organization decides who may be on it, under a documented authorization, and loads their images.
- Capture and compare. Cameras at a deployment point send faces to a recognition model, which scores each face against the watchlist. Faces below the threshold are deleted straight away.
- Alert a trained person. A score above the threshold produces an alert with both images, and a trained operator or officer verifies it before anyone is approached.
- Act and record. The operator decides what to do under a written policy, and the alert, the decision and the outcome are logged.
- Review. The organization tracks false alerts and checks results by demographic group, and a regulator or auditor reviews the governance.
- Audience
- Back office
- Autonomy
- Assist
- Adoption
- Early adopters
What is it worth?
Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.
No public deployment has published a number for the KPIs of this use case yet.
Value drivers: Risk and loss reduction, Speed and cycle time.
Who already uses it?
5 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.
Essex Police
United Kingdom · Government and public sector · 2025
Essex Police used live facial recognition cameras to match passers by against watchlists. The force paused its live deployments after identifying potential accuracy and bias risks. The Information Commissioner's Office audited the force in November 2025 and noted the pause in its March 2026 commentary, and says it continues to work with the force to make sure these risks are addressed.
No outcome disclosed.
South Wales Police
United Kingdom · Government and public sector · 2025
South Wales Police and Gwent Police use live facial recognition, where cameras match the faces of passers by to a watchlist in real time, and retrospective facial recognition, where images from investigations are compared with a database. In August 2025 the UK Information Commissioner's Office published the executive summary of its consensual data protection audit of both forces and gave a high level of assurance for both uses, with recommendations in each. The ICO audit covers governance and use of personal information, not a technical test of the software. It reports that the forces kept human oversight so that no decisions are solely automated.
No outcome disclosed.
Foodstuffs North Island
New Zealand · Retail and ecommerce · 2024
Foodstuffs North Island ran a live facial recognition trial in 25 of its supermarkets from February to September 2024 to identify people previously involved in seriously harmful behavior. Staff verified every match alert with two trained people before acting.
New Zealand's Privacy Commissioner ran an inquiry and found the trial complied with the Privacy Act 2020, with a high level of intrusion reduced by safeguards: the alert threshold was raised from 90 percent to 92.5 percent likelihood of a match, access was restricted and images could not be used for training.
The Commissioner found the technology effective at reducing harmful behavior, especially serious violent incidents, but said the software was developed overseas and not trained on the New Zealand population, so bias against Maori and Pacific people could not be ruled out, and asked for monitoring of the effect of skin tone.
- Interactions handled: 226 million, 8 February to 7 September 2024, 25 supermarkets
"During the trial, 225,972,004 faces were scanned (includes multiple scans of the same person), with 99.999% of these deleted within one minute, and there were 1742 alerts of which 1208 were confirmed matches."
Claimed by: regulator or auditor
Rite Aid
United States · Retail and ecommerce · 2020
From 2012 to 2020 Rite Aid used artificial intelligence based facial recognition in hundreds of stores to identify customers who may have shoplifted.
The US Federal Trade Commission alleged that the system generated thousands of false positive matches, that it was more likely to flag people in stores in plurality Black and Asian communities, and that Rite Aid never tested the accuracy before deployment, did not monitor false positive rates, used low quality images and did not train staff to treat matches as possibly wrong. Employees followed, searched and ejected customers and called police on the strength of alerts.
To settle, Rite Aid agreed to a five year ban on facial recognition for surveillance and to delete the images. The FTC statement describes allegations in a complaint that accompanied a settlement.
No outcome disclosed.
Kmart Australia
Australia · Retail and ecommerce · 2022
From June 2020 Kmart ran a facial recognition system in 28 stores to detect refund fraud and to identify people suspected of earlier theft or refund fraud, capturing the faces of everyone entering those stores. It ran until the Australian privacy regulator opened an investigation in July 2022. According to iTnews and the law firm Clayton Utz, the Privacy Commissioner determined in 2025 that collecting this biometric information without consent breached the Privacy Act, because less privacy intrusive methods existed, and directed Kmart to stop using the system in its existing form. Wesfarmers said Kmart was disappointed and reviewing its options to appeal.
No outcome disclosed.
How do you implement it?
A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.
Data you need
- A watchlist with a documented lawful basis, an owner and a review date for every entry
- Cameras with image quality that meets the model's requirements at the deployment points
- A test set of the local population to measure false matches by demographic group
Systems to integrate
- Camera and video management systems
- Case or incident management system for logging alerts and outcomes
- Operator verification interface with both images side by side
- Retention and deletion controls for non matching faces
Complexity: High
The matching model is bought from a specialist vendor. The build is the legal basis and impact assessment, the watchlist governance, the cameras and their placement, the integration with the verification and incident systems, and the bias and accuracy testing on the local population.
- 1
Establish the legal basis and assess the impact
Write down why facial recognition is necessary and proportionate, complete the data protection impact assessment and consult the regulator where the law expects it, before buying anything.
- 2
Test accuracy and bias on your population
Measure false matches by age, skin tone and sex on images that look like your own cameras produce, and set the alert threshold from the result, not from the vendor default.
- 3
Run a limited, announced trial
Start at a few locations with visible signage, count scans, alerts, confirmed matches and false alerts, and publish the findings or share them with the regulator.
- 4
Review and widen only on the results
Expand when false alerts and demographic gaps are within the limits you set, and keep an independent audit on a schedule.
Guardrails
- A trained person verifies every alert before any action, and no decision is solely automated
- Faces that do not match are deleted straight away and never used to train the model
- A watchlist limited to named categories, with an authorization and a review date per entry
- Visible notices at every deployment point
- Regular bias testing, with the threshold raised when the gap persists
KPIs to instrument
- Confirmed matches and false alerts per 100,000 faces scanned
- False alert rate by demographic group
- Time from alert to verification
- Complaints and challenges, with the outcomes
Human in the loop
Trained operators confirm or discard every alert and decide the action. A senior officer or manager authorizes each deployment, and an independent person reviews false alerts and complaints.
Common failure modes
- Unreliable alerts for some groups
- A model trained on a different population can misidentify some groups more often. The New Zealand Privacy Commissioner noted that the Foodstuffs software was not trained on the New Zealand population, so bias could not be ruled out. Test locally and monitor by group.
- Staff act on an unverified alert
- Following, searching or ejecting a person on an unverified alert is what the FTC alleged at Rite Aid. Require verification by a trained person and a written action policy.
- Watchlist creep
- The list grows beyond serious harm or wanted offenders. Give every entry an owner and a review date, and remove stale entries.
What are the risks and rules?
EU AI Act
High risk
Two provisions apply. Article 5(1)(h) prohibits the use of real time remote biometric identification in publicly accessible spaces for law enforcement, a prohibition that applies since 2 February 2025. The only exceptions are a targeted search for specific victims of abduction, trafficking or sexual exploitation and for missing persons, the prevention of a specific, substantial and imminent threat to life or physical safety or of a terrorist attack, and locating a suspect of a serious crime listed in Annex II.
Each use needs prior authorization by a judicial or independent administrative authority (Article 5(3); in a duly justified urgency it must be requested within 24 hours) and a national law that allows it (Article 5(5)). Live police watchlist matching in the EU is therefore prohibited outside those exceptions.
Every remote biometric identification system that is not prohibited, including retrospective matching by police and live matching by a retailer, is high risk under Annex III point 1(a); Article 26(10) adds an authorization duty for retrospective police searches. The schema has no prohibited tier, so the page is recorded as high. Faces are biometric data, so a retailer also needs a condition under GDPR Article 9.
Guidance
- Regulation (EU) 2024/1689 (AI Act), Article 5(1)(h) and Annex III point 1 (European Union, Europe). Article 5(1)(h) prohibits real time remote biometric identification in publicly accessible spaces for law enforcement outside narrow exceptions; Annex III point 1(a) makes other remote biometric identification high risk.
- ICO audit of South Wales Police and Gwent Police, executive summary (Information Commissioner's Office, Europe). Shows what a regulator checked for live and retrospective facial recognition: necessity and proportionality, fairness and accuracy, and human oversight.
- Supermarket trial of FRT, Privacy Commissioner inquiry results (Office of the Privacy Commissioner (New Zealand), Asia Pacific). A regulator inquiry into a live facial recognition trial in 25 supermarkets, with the safeguards the Commissioner relied on and the improvements it asked for before any permanent use.
Controls to put in place
- Data protection impact assessment completed and reviewed before each new deployment
- Written authorization per deployment with the legal basis, the watchlist and the location
- Audit trail of every alert, verification and action
- Independent audit or inquiry on a schedule
When it went wrong elsewhere
- Rite Aid banned from using AI facial recognition after FTC says retailer deployed technology without reasonable safeguards. The FTC alleged that the system, used in hundreds of stores from 2012 to 2020, produced thousands of false positive matches, more often in plurality Black and Asian communities, and that staff followed, searched and ejected customers on the strength of alerts. Rite Aid accepted a five year ban.
Frequently asked questions
- Have regulators accepted any facial recognition deployments?
Some have reviewed specific deployments and found them compliant, with conditions. The UK Information Commissioner's Office gave South Wales Police and Gwent Police a high level of assurance in its 2025 data protection audit, and the New Zealand Privacy Commissioner found the Foodstuffs North Island supermarket trial complied with the Privacy Act 2020. Both findings depend on human review of alerts and written safeguards, and neither is a general approval. In the EU, live police use is prohibited outside narrow exceptions.
- What went wrong at Rite Aid?
The US Federal Trade Commission alleged that Rite Aid used the technology without reasonable safeguards, that it produced thousands of false positive matches and that staff acted on them against customers. Rite Aid agreed to stop using facial recognition for surveillance for five years.
- Is facial recognition for watchlists allowed under the EU AI Act?
Real time remote biometric identification in publicly accessible spaces for law enforcement is prohibited by Article 5(1)(h), except for a few listed purposes with prior judicial or independent authorization and a national law that allows it. Other remote biometric identification, such as retrospective searches or a retailer's system, is high risk under Annex III point 1(a). The UK deployments on this page fall under UK law, not the EU AI Act.
How to cite this page
Blits.ai AI Use Case Library, "AI facial recognition for matching faces against watchlists in policing and retail", last verified 12 November 2026, https://www.blits.ai/ai-use-cases/facial-recognition-watchlist-matching. Licensed under CC BY 4.0. Method: how we verify use cases.
Changelog
- 12 November 2026: First published