Facial recognition CCTVwhat AI cameras actually do, and what the law asks of a shop
AI is printed on a lot of camera boxes now. Most of it is detection, which tells a person from a cat and finds the clip you need. Face recognition is a different thing with different rules. Here is what each one does, which UniFi cameras can recognise faces, and what UK law asks before a shop switches it on.
The quick answer
Most "AI" in CCTV is detection: the camera works out that a person, a vehicle, an animal or a face is in view, so alerts and searches stop wasting your time. Facial recognition is different. It matches a face against stored faces to say who someone is, and in the UK that makes it biometric special category data. A shop can only do it with an Article 9 condition, a written policy document and an impact assessment done first. For most shops, sharp face shots at the door and till plus fast search do the job without any of that.
What this page covers
- Face detection, face recognition and the other AI detections, told apart.
- Which UniFi cameras recognise faces, and what needs an AI Port or AI Key.
- Why recognising a face is special category data under UK GDPR.
- What a shop must have in place before matching a single face.
- Shared watchlist services and the ICO's Facewatch decision.
- The Bridges case, the Home Office's new framework and Scotland.
- Homes, and where the household exemption stops.
Five things an AI camera can do
The legal weight column is the one to read. Four rows are ordinary CCTV. One is not.
| What the camera does | In plain terms | In a shop | Legal weight |
|---|---|---|---|
| Motion detection | Notices pixels changing | A hedge moving sets it off as easily as a person | Ordinary CCTV rules |
| Person, vehicle and animal detection | Works out what kind of thing moved | Alert for a person in the stock room after closing, not for a cat | Ordinary CCTV rules |
| Face detection | Spots that a face is in view and picks out the moment | Jump straight to every clear face at the door between 2pm and 3pm | Ordinary CCTV rules, as long as nothing is matched |
| Smart search | Finds footage by what was detected | Every person near the till, or every red car, across all cameras | Ordinary CCTV rules |
| Face recognition | Compares a face against stored faces to say who it is | Alert when a face on your list walks in | Biometric special category data: Article 9 condition, policy document, DPIA |
Motion detection
- In plain terms
- Notices pixels changing
- In a shop
- A hedge moving sets it off as easily as a person
- Legal weight
- Ordinary CCTV rules
Person, vehicle and animal detection
- In plain terms
- Works out what kind of thing moved
- In a shop
- Alert for a person in the stock room after closing, not for a cat
- Legal weight
- Ordinary CCTV rules
Face detection
- In plain terms
- Spots that a face is in view and picks out the moment
- In a shop
- Jump straight to every clear face at the door between 2pm and 3pm
- Legal weight
- Ordinary CCTV rules, as long as nothing is matched
Smart search
- In plain terms
- Finds footage by what was detected
- In a shop
- Every person near the till, or every red car, across all cameras
- Legal weight
- Ordinary CCTV rules
Face recognition
- In plain terms
- Compares a face against stored faces to say who it is
- In a shop
- Alert when a face on your list walks in
- Legal weight
- Biometric special category data: Article 9 condition, policy document, DPIA
Every row still needs the normal CCTV basics: a reason for the camera, signs, sensible retention and secure storage. Only the last row turns the footage into biometric data.
By the Doberman install team · UniFi CCTV designers & installers, England · Last reviewed November 2026
By the Doberman install team
UniFi CCTV designers & installers, England
Last reviewed November 2026
What "AI" in a CCTV camera actually does
An AI security camera analyses its own picture. Older cameras only noticed change: a block of pixels moved, so they recorded and pinged your phone. That is why a basic system alerts you to every car headlight and every bush in the wind. An AI camera runs a trained model over the image and decides what moved.
Five jobs get sold under the same "AI CCTV" label, and they are worth keeping apart:
- Smart detection. Person, vehicle, animal or package. The camera tags the event, so you can be told about a person at the back door after closing and ignore a fox.
- Face detection. The camera finds a face in the frame and marks the moment, so you can scroll to the clear shot instead of scrubbing through an hour. It knows there is a face. It does not know whose.
- Face recognition. The system compares the face with faces it has stored and says "this is person X" or "this face is unknown". This is the one with the legal weight.
- Licence plate recognition. Reading number plates and matching them to a list. Our guide to ANPR cameras covers it in full.
- Smart search. Searching recorded footage by what was detected, across every camera at once, instead of by time alone.
The line that matters is between the second and third bullets. Face detection gives you a usable picture of a face. Face recognition turns that picture into a biometric identifier and checks it against a list. The ICO's biometric guidance says plainly that recognition systems "can therefore never be certain about an individual's identity", and that you should not treat a match as "an objective fact about someone's identity". A match is a suggestion for a person to check.
What UniFi Protect does with faces
On Ubiquiti UniFi Protect, the systems we install, face recognition is a real feature, and which hardware does it is clearly set out. G6 and AI series cameras have it built in, apart from the 360 models. Older cameras need an accessory. The newest features, such as face enhancement and natural-language search, need an AI Key on the system whichever camera you have.
| UniFi camera | Person and vehicle detection | Face recognition | Newest AI features |
|---|---|---|---|
| G6 and AI series (not 360 models) | Built in | Built in | AI Key for face enhancement, natural-language search and AI summaries |
| G5 and G4 | Built in | With an AI Port (real time) or an AI Key (queued, can lag) | AI Key for the same extra features |
| G3 | With an AI Port | With an AI Port | AI Key and AI Port |
| Third-party ONVIF cameras | With an AI Port | With an AI Port | AI Key and AI Port |
| AI LPR | Built in | Not listed | Dedicated to number plates |
G6 and AI series (not 360 models)
- Person and vehicle detection
- Built in
- Face recognition
- Built in
- Newest AI features
- AI Key for face enhancement, natural-language search and AI summaries
G5 and G4
- Person and vehicle detection
- Built in
- Face recognition
- With an AI Port (real time) or an AI Key (queued, can lag)
- Newest AI features
- AI Key for the same extra features
G3
- Person and vehicle detection
- With an AI Port
- Face recognition
- With an AI Port
- Newest AI features
- AI Key and AI Port
Third-party ONVIF cameras
- Person and vehicle detection
- With an AI Port
- Face recognition
- With an AI Port
- Newest AI features
- AI Key and AI Port
AI LPR
- Person and vehicle detection
- Built in
- Face recognition
- Not listed
- Newest AI features
- Dedicated to number plates
From Ubiquiti's AI detections, AI Key and AI Port help articles and its tech specs pages, read 8 November 2026. The AI Port does not work with AI series cameras, which do not need it.
Where the analysis happens. Ubiquiti says its AI processing "takes place entirely on-premises", on the camera or on an AI Port or AI Key, "so your video never needs to be sent to the cloud for analysis". It also says Protect stores video and AI events locally and that "We cannot view, store, or access your video streams at any point." For a business, that means no outside company is doing the matching on your footage. It does not reduce your legal duties. You are still the one responsible for what the system does.
Known and unknown faces. Protect groups the faces it sees. Name one and it becomes a known person; faces can be reassigned or deleted in Protect's Find Anything search. Alarm Manager can then trigger on "Known/unknown faces, license plates, persons of interest", with actions such as a push alert or switching on a floodlight. In a shop, that alarm is exactly the watchlist the law is most concerned with.
The AI Key's own limits. Ubiquiti is candid about these. It says the AI Key's models "might occasionally return inaccurate results", that it holds a queue of 200 events and discards the excess, that a full queue "can delay processing up to 15 minutes", and that natural-language search is English only and covers only detections the AI Key has processed. Fine for finding "person walking a dog" in last night's footage. Not something to build an instant ban-list alert on.
Settings that help you comply. Protect has Privacy Zones, which Ubiquiti says are blacked out and not recorded, often "to comply with local laws and regulations", and Exclusion Zones that "will not generate AI-powered detections". Use them on the neighbour's window, the pavement beyond your door or a staff break area. Retention is set per camera, which matters for the next section. For the rest of the app, see UniFi Protect explained.
Why recognising a face is special category data
Ordinary CCTV footage is personal data. A face run through recognition is biometric data, and UK GDPR puts that in its most protected class. Article 4 defines biometric data as personal data from "specific technical processing" of a person's physical characteristics "which allow or confirm the unique identification of that natural person, such as facial images". Article 9 then says processing "biometric data for the purpose of uniquely identifying a natural person" is prohibited unless an exception applies.
The point most people miss is that this covers everyone the system scans, not only the people it matches. The ICO's biometric guidance says that if your purpose is to identify people, "you are processing special category biometric data from the moment you collect the biometric data", and that this holds "even if: you do not find a match". A camera on the shop door that checks every face against a list of ten is processing the biometric data of every customer who walks in.
To do that lawfully, a business needs three things in place before the system starts:
- A lawful basis under Article 6. Since 5 February 2026 UK GDPR has a list of "recognised legitimate interests", and one of them is processing necessary for "detecting, investigating or preventing crime". That covers Article 6 only. Article 9, as amended, still needs its own condition on top.
- An Article 9 condition. Explicit consent is the first one listed, and it rarely works here. The ICO says that if you are systematically monitoring a public space "it is highly unlikely that consent is appropriate", and you "must not assume that someone walking into the field of view" has consented. That leaves substantial public interest, which for a shop usually means the condition in paragraph 10 of Schedule 1 to the Data Protection Act 2018.
- An appropriate policy document. Paragraph 10 only works if the business has written one. It explains how you comply with the data protection principles and how long you keep the data before it is erased. The ICO says it must be in place "at the time your processing starts".
Paragraph 10, headed "Preventing etc unlawful acts", is met only if the processing:
- is necessary for the prevention, investigation or detection of an unlawful act;
- must be done without the person's consent so as not to prejudice that purpose; and
- is necessary for reasons of substantial public interest.
The ICO adds that "It is not enough for you to make a vague or generic public interest argument", that you must explain "how the processing is likely to be effective", and that using biometric data to detect or prevent crime means you are likely to be processing criminal offence information as well. If no condition fits, its answer is blunt: you "must not use a biometric recognition system."
One date to keep an eye on. Every ICO page used here carries a banner saying it is "under review" because of the Data (Use and Access) Act 2025. The Article 9 ban itself has not been lifted. Check the ICO's current version before you rely on any detail.
What a shop needs before matching one face
The ICO's biometric guidance, its facial recognition guidance and the Commissioner's 2021 Opinion on live facial recognition in public places set out the work a shop takes on. The Opinion says that where faces are collected automatically and indiscriminately in public places, "there is a high bar for its use to be lawful".
- A DPIA, done first. "You must complete a DPIA before you use a biometric recognition system." Treat it as "a living document" reviewed before every deployment.
- Necessity, with the alternatives written down. Explain "why you have ruled out less intrusive options" and "how you have measured its effectiveness". For shopping centres the ICO says it may be harder to justify "processing images of large numbers of individuals to only identify a few".
- Accuracy and bias checked. The system must be "sufficiently accurate for your purposes", and you "should test for and mitigate biases" so they do not lead to discrimination.
- A person checks every match. Human input must be "active, and has a meaningful influence on the outcomes", and you must be able to tell false matches from true ones.
- Watchlist rules. Strictly limit who goes on it, keep images only as long as necessary, and use only images that are "lawfully acquired and accurate". The Opinion says watchlists of people suspected of minor offences "are less likely to satisfy the key legal tests", and images taken from social media raise problems of their own. Keep the list as small as you can.
- Signs that say facial recognition. "Adapting standard CCTV signage is likely to be insufficient." Signs should say facial recognition is in use and why, that biometric data is processed, and how to find out more and use your rights. They have to be read before the camera captures someone, and it goes in your privacy notice too.
- Retention you can defend. UK law sets no fixed period, and a long default on the recorder is no reason on its own to keep anything. Write down how long non-matches, matches and watchlist entries are kept and why.
- Know who the controller is. If an outside service runs the matching or the watchlist, you need to know if you are a controller or a joint controller, and you need a written contract with any processor.
- A way to challenge it. The ICO says this may mean giving people a way to challenge an outcome they think is wrong.
None of that is a camera setting. It is a written project with an owner, and it needs staff trained in what to do when an alert fires. Our guide to UK CCTV regulations covers the ordinary CCTV duties that sit underneath all of this.
Want an AI camera system that stays on the right side of the line?
Tell us about the shop: the doors, the till, the stock room and what keeps going missing. We will design UniFi cameras with person and vehicle detection and clear face shots where the police need them.
Shared watchlists and the ICO's Facewatch decision
The other route into facial recognition is to subscribe to a service that runs the system and the watchlist for you. Facewatch is one such service, used by a number of UK retailers. The ICO described its system this way: it "scans people's faces in real time as they enter a store and alerts if a 'subject of interest' has entered." The shops that use it are its subscribers.
The ICO looked at Facewatch and published its view on 31 March 2023. It said it agreed "the company had a legitimate interest in using people's personal data", but "identified various areas of concern". Facewatch made changes: "focusing on repeat offenders or individuals committing significant offences", appointing a Data Protection Officer, and making sure people classed as vulnerable do not become a "subject of interest". The ICO then concluded "no further regulatory action is required".
Three things in that decision matter for any shop thinking of signing up:
- The ICO said it is "not a blanket approval of Facewatch, nor of LFR use" and that "Each new application must be considered on its own merits". Your store is a new application.
- Its closure letter, dated 28 March 2023, said the processing as first examined "failed to balance the legitimate interest of Facewatch and their subscribers against the rights and freedoms of individuals". The subscribers were part of that balance, which is why the controller question above matters.
- The letter also said the purpose of preventing crime is a legitimate interest and that "Legal gateways in the data protection legislation allow biometric data to be processed for this purpose." So it can be lawful, done properly.
The weak point in practice is the person reading the alert. In August 2026 The Register reported staff at a Sainsbury's store asking the wrong shopper to leave after an alert. Sainsbury's said it was caused by human error, not the technology, and Facewatch said a correct alert had been handled wrongly in store. That is the human review duty in practice: the match is the start of a check, not the end of one.
The ICO's blog has since been taken off its live site; the copy kept by the National Archives is linked under About this page. If you are thinking of sharing photos of shoplifters some other way, such as with neighbouring shops or online, read whether shops can share photos of shoplifters first.
What the courts and government have said
Bridges v South Wales Police (2020). The leading UK case is about police, not shops. On 11 August 2020 the Court of Appeal found South Wales Police's use of live facial recognition unlawful on three grounds. It said "It is not clear who can be placed on the watchlist nor is it clear that there are any criteria for determining where AFR can be deployed". It found the impact assessment "failed properly to assess the risks to the rights and freedoms of data subjects". And it found the force had never satisfied itself that the software "does not have an unacceptable bias on grounds of race or sex."
The specific duties in that case, the police data protection rules and the public sector equality duty, do not bind a private shop. The three failures still map onto the shop checklist above: who goes on the list and why, an impact assessment that deals with the real risks, and checking the software for bias.
The Home Office framework. The Home Office consulted on a legal framework for facial recognition in law enforcement from 4 December 2025 to 12 February 2026, and published its response on 4 November 2026. It will bring in "a new statutory framework for use of biometrics, facial recognition, and similar technologies by law enforcement and other public bodies", with a new independent oversight body. The King's Speech on 13 May 2026 put that framework in a Police Reform Bill, most of which applies to England and Wales.
For shops, the response says "Private-sector use is already subject to existing data protection law", and commits to "a Code of Practice for private-sector use for the prevention and detection of crime". No date has been set for that code. Until it arrives, the rules for a shop are the UK GDPR and ICO rules above.
Scotland. The Data Protection Act 2018 extends to Scotland, so a shop in Scotland works to the same UK GDPR, the same Schedule 1 conditions and the same ICO as one in England or Wales. Scotland has its own Biometrics Commissioner, but its functions extend to Police Scotland, the Scottish Police Authority and the Police Investigations and Review Commissioner, and its Code of Practice, in force since 16 November 2022, covers biometric data used for criminal justice and police purposes. The Bridges judgment is from the Court of Appeal of England and Wales.
Facial recognition at home and the household exemption
UK GDPR does not apply to processing "by an individual in the course of a purely personal or household activity". That is why a homeowner whose cameras only cover their own house and garden has few formal duties. Recognising your own family as they come up your own drive sits on that safe ground.
The exemption stops at the boundary. In the Rynes case in 2014, the EU's top court said a home camera that "covers, even partially, a public space" is not a purely personal or household activity; that was before Brexit, but it is where the UK line comes from. In Fairhurst v Woodard in 2021, an Oxford county court judge treated a homeowner whose cameras collected data beyond his boundary as having to justify it under data protection law, and found that his driveway camera "is not lawful" and that he had "breached the provisions of the DPA 2018 and the UK GDPR".
Put the two together. A doorbell or driveway camera that takes in the pavement already brings data protection law into play. Switch on face recognition and every passer-by becomes biometric processing, which is very hard for a household to justify. Our advice for a home: use person and package detection, aim the cameras at your own property, and leave recognition of strangers off. The everyday version of this is in neighbours' CCTV and the law.
What most shops should do instead
For most shops, a face the police can use, and footage you can find in a minute, does more for less legal risk than a watchlist. Here is the version that needs no biometric processing:
- An identify-quality camera on the door. Home Office guidance for small businesses says at least one camera must give "identifiable quality images of everyone entering your premises", and that the main doorway is usually the best place. Its identify level means the person fills the screen height, enough "to identify an individual beyond reasonable doubt".
- A second one at the till. The same guidance strongly recommends a second camera of the same quality near the till.
- Detection everywhere else. Person detection on the aisles, stock room and back door, with alerts outside opening hours.
- Search and export. Smart search to find every person at the till in a time window, then a clean export the police can play. Our guide to exporting CCTV footage for the police walks through it.
The same guidance warns "don't install a camera for one purpose and expect to use it for another". A wide camera over the shop floor will not give you a face at the door, whatever AI it runs. Placement and lighting decide that. For camera counts, see how many CCTV cameras a small shop needs, and for the full shop plan, our guide to CCTV for shops.
If a business decides it does need recognition, the UniFi hardware above can do it, and the right order is the legal work first, then the switch. On a UniFi system that means the DPIA, the policy document, the watchlist rules and the new signs are done before face recognition is turned on for a single camera.
Want clear faces at the door and till, and footage the police can use?
Tell us how the shop is laid out and where the trouble starts. We will plan UniFi cameras for identify-quality faces at the entrance and till, AI detection on the rest, and a recorder that keeps the evidence on site.
About this page
Who it is for
Shop owners and managers, and homeowners, weighing up an AI camera system and wondering whether to switch face recognition on. Reviewed 8 November 2026.
Sources
The law is from UK GDPR Article 9, Article 4, Annex 1 and Schedule 1 paragraph 10 of the Data Protection Act 2018 on legislation.gov.uk. ICO guidance is from its biometric recognition guidance, its facial recognition and surveillance guidance, the Commissioner's Opinion on live facial recognition (18 June 2021), the ICO blog on Facewatch (31 March 2023, National Archives copy) and its closure letter. The court cases are R (Bridges) v Chief Constable of South Wales Police, Fairhurst v Woodard and Rynes. Government policy is from the Home Office's facial recognition consultation and response and Scotland's position from the Scottish Biometrics Commissioner. The shop camera standard is the Home Office's CCTV guidance for small businesses. UniFi details are from Ubiquiti's AI detections and facial recognition, AI Key and AI Port articles and its tech specs pages.
What we have deliberately not said
We have not given accuracy or false-match percentages for any system: the ICO's point is that a match is never certain, whatever the figure. We have not quoted prices for UniFi hardware. And this page is a plain-English guide to the rules, not legal advice for your business: a recognition system needs its own DPIA, and the ICO's guidance is under review.
Common questions
Is facial recognition CCTV legal in the UK?
It can be, but it is the most heavily regulated thing a camera can do. Using a face to identify a specific person creates biometric data, which UK GDPR Article 9 treats as special category data and prohibits unless you have a lawful basis plus one of its listed conditions. For a shop that usually means the crime prevention condition in the Data Protection Act 2018, a written appropriate policy document and a data protection impact assessment completed before the system goes live. The ICO describes the bar for live facial recognition in public places as high.
What is the difference between face detection and face recognition?
Face detection spots that a face is in the picture, so the camera can flag the moment or keep a clear shot of it. Face recognition goes a step further and compares that face against faces already stored, to say who it is. Detection on its own does not identify anyone. Recognition does, and the ICO says a business using a recognition system is processing special category biometric data from the moment it collects it, even when no match is found.
Can a shop use facial recognition to catch shoplifters?
Some UK retailers do, so it is possible, but each has to justify it. The shop has to show that matching faces is necessary to prevent or detect crime, that asking for consent would defeat the purpose and that there is a substantial public interest. It needs an appropriate policy document, a DPIA before it starts, strict rules on who goes on the watchlist, signs that say facial recognition is in use, and a person checking every alert before anyone acts on it.
Do I need a DPIA for facial recognition?
Yes. The ICO says you must complete a data protection impact assessment before you use a biometric recognition system, and its surveillance guidance calls the DPIA a living document to review before every deployment. It should explain why you need facial recognition, why less intrusive options will not do the job, how you will measure whether it works, and how you will handle accuracy, bias and false matches.
What is an AI security camera?
A camera, or a camera and recorder together, that analyses the picture instead of only reacting to movement. It can tell a person from a car, an animal or a swaying hedge, so alerts fire on what matters and you can search footage by what happened. Some can also read number plates or recognise faces. On Ubiquiti UniFi Protect, that analysis runs on the camera or on an add-on on your own site, not in the cloud.
Which UniFi cameras have face recognition?
According to Ubiquiti, the G6 and AI series cameras have face recognition built in, except the 360 models. G5 and G4 cameras detect people and vehicles on their own but need an AI Port accessory for real-time face recognition; an AI Key can also add it, with processing delays when its queue is busy. Older G3 cameras and third-party ONVIF cameras need an AI Port. The AI LPR camera is built for number plates and does not list face recognition.
Does UniFi face recognition send faces to the cloud?
No. Ubiquiti says its AI processing takes place entirely on-premises, on the camera itself or on an AI Port or AI Key, so video never needs to go to the cloud for analysis. It says Protect stores video and AI events locally and that Ubiquiti cannot view, store or access your video streams. That keeps the data on your own recorder, but it does not change your legal duties: you are still the business responsible for it.
Can I use facial recognition cameras at home?
Purely personal or household use falls outside UK GDPR, so recognising your own family on cameras that only cover your own property sits on safe ground. Once a camera takes in the street, a neighbour's garden or a shared path, data protection law applies to you, as the court found in Fairhurst v Woodard. Recognising the faces of passers-by would then be special category processing, which is very hard for a household to justify.
What did the ICO say about Facewatch?
In March 2023 the ICO said it agreed Facewatch had a legitimate interest in using personal data to prevent crime, but found areas of concern. Facewatch changed its system to focus on repeat offenders or significant offences, appointed a Data Protection Officer and stopped vulnerable people becoming subjects of interest. The ICO then took no further regulatory action, and said the decision was not a blanket approval of Facewatch or of live facial recognition.
Want AI cameras that catch the face without the legal headache?
Answer a few quick questions and we will quote from your answers. If anything needs a look in person, we arrange a survey first.
