The Centre for Retail research has produced an overview of crime statistics in the UK which show a continued increase in the level and type of crime facing the retail sector.

There are several official and unofficial surveys of retail crime in the UK, including the Home Office Crime Against Business 2018, the British Retail Consortium’s Retail Crime Survey 2018, the Association of Convenience Stores’ The Crime Report 2019, as well as Sensormatic’s Global Shrink Index and Checkpoint’s Global Retail Theft Barometer. Each one has a different methodology and covers a slightly different purpose.

We have consolidated this information into a simple info-graphics with the focus on the major crime and violence issues facing UK retailers

Download the info-graphics guide:

FaceWatch_Infographic_A5 v6-2

SBT are one of the leading providers of technology solutions to the retail sector in Spain and believe that facial recognition as a retail crime deterrent, as in the UK, is an important tool to improve the retail experience for both customers and staff.  With an international customer base and a focus on customer behaviour analysis, shopping experience and loss prevention the adoption of Facewatch the leading AFR system into SBT’s portfolio was seen as a natural step.

Jose Alvarez Abad, VP Strategic & Sales, SBT said:

“We have been watching the development of AFR for the retail sector with interest as Spain unfortunately experiences the same type of issues that plague the UK retail sector. As a specialist provider of retail technology, it became apparent that with the current European data laws Facewatch was the product that met all our requirements and gives us the opportunity to run trials with our key clients immediately.”

Nick Fisher, CEO, Facewatch Ltd

“Facewatch is fast becoming the standard for retail sector facial recognition in the UK and this has led to huge interest across Europe and the World. In 2016 we launched our first version of Facewatch facial recognition in Brazil and using our real world experience and early adoption of the GDPR privacy principles we have created an incredibly simple and effective system that is fully GDPR compliant.  The opportunity to work with SBT and appoint them as our distributor in Spain is great news for us, as having the opportunity to work with a well-established technology focused partner like SBT allows us to scale in other countries in the knowledge the training, implementation and GDPR compliance will be expertly managed.


SBT company background

SBT is a Spanish company specialising in the analysis, design and implement high technology solutions to the retail market. SBT’s clients include Pepe Jeans, Tommy Hilfiger and Pikolinos.

Established in 2017 the business is formed by a group of professionals with a broad experience with the focus on 4 business areas: In-store customers and product behaviour, operations and shopping experience,  security and loss prevention analytics and general retail services.

Facewatch company background:

Facewatch have been providing crime prevention solutions to the retail industry for over 10 years. The business was started by Simon Gordon owner of London’s oldest wine bar on the Embankment in London. The Wine bar was a target for pick pockets and bag thieves and he wanted to provide a relaxed and safe environment for his customers. Being technology minded and working with the local police he launched the first ever online crime reporting system including CCTV footage. This led to the launch of the first facial recognition solution in 2017, enabling retailers to deter habitual criminals who were shoplifting, abusing staff or causing criminal damage.

Today the Facewatch system provides a GDPR compliant solution that is easy to install, can be used and managed by small stores and is scalable for use by large retail groups due to its unique cloud-based servers and using Intel® NUC mini PCs. Data is managed securely by Facewatch. Facewatch doesn’t store information about the general public, just those for whom their retailer subscribers have uploaded confirmed evidence of criminal activity. If a facial image is not matched to a relevant watch list the algorithmic data is instantly deleted.

Figure 1. Facewatch matches faces against known offenders within seconds of them entering a business

Facewatch technology

Facewatch solution overview:

Facewatch uses the software-as-a-service technology model, making advanced facial recognition affordable for even small businesses. The company’s database of Subjects of Interest (SOI’s) is stored securely in the cloud. It’s a centralised, managed database of biometric data corresponding to the faces of people who are reasonably suspected of having shoplifted or committed other crimes at businesses that subscribe to the service (Figure 1).

The hardware to run Facewatch is simple to deploy. It includes a standard HD CCTV camera and Intel® NUC, a mini-PC that is only 4×4 inches in size and consumes very little power. Its performance enables it to play and record video at 4K Ultra HD clarity, making it ideal for a facial recognition system. The cameras—placed at store entrances—send an image to an on-site NUC loaded with software that converts the image to an algorithm. The algorithm is compared to those in the Facewatch relevant database for that property and if there is a match an alert—along with an accuracy reading—is sent to the retailer’s smartphone or other device, warning it that a known criminal on the database has entered its business.

To add a shoplifter to the database takes only six key presses and about 20 seconds, making it easy for store or security staff, and it doesn’t interfere with their normal duties. “They simply follow a dropdown menu, the time and date are automated, tick the box, the whole thing’s designed to be simple but highly secure and includes a confirmatory legal statement to ensure that the information is accurate.” said Nick Fisher, CEO of Facewatch.The solution does not retain any personal data on anyone not on the database. “If no match is discovered, the image is deleted in 0.3 seconds” Fisher said, “and the entire process—from the moment a known shoplifter comes through the door, to the instant the retailer gets an alert—takes less than two seconds.”

For more information:

Smart Business Technologies

Jose Alvarez Abad

VP Strategic & Sales

[email protected]

(+34) 629 844 266

Pº de la Castellana 153

28046 Madrid – Spain   



Stuart Greenfield


T: 07788 662697

E: [email protected]



By Charles Hymas

Home Affairs Editor (The Daily Telegraph)

Please accept statistics, marketing cookies to watch this video.

When you enter Paul Wilks’ supermarket in Aylesbury, a facial recognition camera by the door snaps your image and then checks it against a “watchlist” of people previously caught shoplifting or abusing staff.

If you are on it, managers receive a “ping” alert on their mobile phones from Facewatch, the private firm that holds the watchlists of suspects, and you will be asked to leave or monitored if you decide you want to walk around the store.

This is not some Big Brother vision of the future but Budgens in Buckinghamshire in Britain 2019.

It is also stark evidence of the way that Artificial Intelligence (AI) technology is spreading without regulation potentially intruding on our personal privacy.

For Mr Wilks, it has been a success. Since he introduced it at his 3,000 square foot store a year ago, he says shoplifting has fallen from ten to two incidents a week. The thousands he has saved has more than paid for the technology.

“As well as stopping people, it’s a deterrent. Shoplifters know about it,” says Mr Wilks, who has a prominent poster warning customers they will be subject to facial recognition. “As retailers, we have to find ways to counteract what is going on.”

As the retail sector loses £700 million a year to thefts, Facewatch gives store owners a “self-help” solution to the reluctance of police to investigate petty shoplifting.

It is the brainchild of businessman Simon Gordon, whose own London wine bar Gordons was plagued by pickpockets. Using AI technology provided by Intel, a US multinational, he has bold ambitions to have 5,000 high-resolution facial recognition cameras in operation across the UK by 2022.

His firm is close to a deal with a major UK supermarket chain and already has cameras being used or trialled in 100 stores, garages and other retail outlets.

The lenses are mounted by entry doors to catch a full clean facial image, which is sent to a computer that extracts biometric information to compare it to faces in a database of suspects.

Facewatch says there must be a 99 per cent match before the alert is sent to store staff and in consultation with the Information Commissioner Elizabeth Denham has introduced privacy safeguards including immediate automatic deletion of images of “innocent” faces.

“When CCTV first came out 25 or 30 years ago, people thought it was the end of the world, Big Brother and 1984,” says Stuart Greenfield, a Facewatch executive. Now there are six million cameras in London. People either think they are not working or are there to stop terrorists. No-one really worries about it. Facial recognition is the same. Facebook, Instagram and the airports are using it. It is here to stay but it has to be regulated. Everything needs to be controlled because every technology can be used negatively.”

And there’s the rub. MPs, experts and watchdogs, like the Information Commissioner Ms Elizabeth Denham and Paul Wiles, the biometrics commissioner, are concerned facial recognition technology is becoming established if not widespread with little public debate or regulatory scrutiny. They point to critical questions yet to be resolved.

When should facial technology surveillance be used, in what circumstances and under what conditions? And should consent be required before it is deployed?

Judges in a test case against its use by South Wales Police ruled taking a biometric image of a face is as intrusive as a fingerprint or DNA swab. More significantly, unlike with fingerprints or a swab, people have no choice about whether, where or when their biometric image is snapped.

South Wales Police are thought to have scanned more than 500,000 faces since first deploying facial recognition cameras during the Champions League Final at Cardiff’s Millennium Stadium in June 2017. The Met Police and Leicestershire police have scanned thousands more in their “trials.”

The test case in South Wales was brought by Ed Bridges, a former LibDem councillor, who noticed the cameras when he went out to buy a lunchtime sandwich. He said taking “biometric information without his consent” was a breach of his privacy when he was acting lawfully in a public place.

The judges, however, ruled use of the technology was “proportionate.” They said it was deployed in an “open and transparent” way, for a limited time to identify particular individuals of “justifiable interest” to the police and with publicity campaigns to alert the public in advance.

However, Mr Wiles, the biometrics commissioner, is not convinced that this test case alone should be taken as sufficient justification for a roll-out of police use of facial recognition. “I am not disagreeing with the South Wales Police judgement. What South Wales Police did was lawful,” he says.

“Some uses of Automated Face Recognition in public places when highly targeted – for example scanning the faces of people going into a football match against watchlists of those known to cause trouble in football matches in the past – that is arguably in the public interest.

“However, scanning everyone walking down the street against a watchlist of people you would like to arrest seems to be a bit more difficult because it gets near mass surveillance. I don’t think in this country we have ever really wanted to see police using mass surveillance. That’s what the Chinese are doing with facial recognition. There are some lines between legitimate use to protect people who have committed crimes against a rather different use. It is not for me to say where the line is. Nor should it be the police who say where it is. That’s the debate we are not having. I feel it is frustrating that ministers are not talking about it. And before we ask Parliament to decide, we need to have a public debate.”

Cases have already emerged where Mr Wiles’s line appears to have been crossed. Last year, the Trafford Centre in Manchester had to stop using live facial recognition cameras after the Surveillance Camera Commissioner intervened. Up to 15 million people were scanned during the operation.

At Sheffield’s Meadowhall shopping centre, some two million people are thought to have been scanned in secret police trials last year, according to campaign group Big Brother Watch.

The privately-owned Kings Cross estate in London has also had to switch off its facial recognition cameras after it became public. It later emerged the Met Police shared images of suspects with the property firm without anyone’s consent or senior officers and mayor’s office apparently knowing.

Liverpool’s World Museum scanned visitors with facial recognition cameras during its exhibition, “China’s First Emperor and the Terracotta Warriors” in 2018, while Millennium Point conference centre in Birmingham – a scene of demonstrations by trade unionists, football fans and anti-racism campaigners – has deployed it “at the request of law enforcement.”

The Daily Telegraph has revealed Waltham Forest council in London has trialled facial recognition cameras on its streets without residents’ consent and even that AI and facial expression technology is being used for the first time in job interviews to identify the best UK candidates.

As its use widens, one key issue is its reliability and accuracy. The success of the technology’s algorithms in matching a face is improving and is good when there is a high-quality image – as at UK passport control – but less effective with CCTV images that do not always give a clear view.

The US National Institute of Standards and Technology which assesses the ability of algorithms to match a new image to a face in a database estimates it has improved 20-fold between 2014 and 2018.

However, it is by no means infallible. In South Wales, police started in 2017 at a Champions League match with a “true positive” rate – where it got an accurate match with a “suspect” on its database – of just three per cent. This rose to 46 per cent when deployed at the Six Nations rugby last year.

Across all events where deployed, there were 2,900 possible matches of suspects generated by the facial recognition system, of which only 144 were confirmed “true positives” by operators; 2,755 were “false positives,” according to the analysis by Cardiff University. Eighteen people were arrested.

The researchers found performance fell as light faded and was less accurate if faces were obscured by clothing, glasses or jewellery. They concluded it should be viewed as “assisted” rather than “automated” facial recognition, as the decision on whether there was a match was a police officer’s.

Professor Peter Fussey, from Essex University, who reviewed the Met Police trials of the technology, said only eight of the 42 “matches” that they saw thrown up by the technology were accurate.

Sixteen were instantly rejected in the van as it was clear they were the “wrong ethnicity, wrong sex or ten years younger,” he said. Four were then lost in the crowd, which left 22 suspects who were then approached by a police officer in the street to show their id or be mobile fingerprinted.

Of these, 14 were inaccurate and just eight were correct.

In the febrile world of facial recognition, how you measure success is a source of debate. It could be argued you have high 90 per cent-plus accuracy given the cameras scan thousands of faces to pick out the 42. Or you can measure it according to the ratio of “false positives” to accurate matches.

On human rights grounds, Professor Fussey said his concern was consent and legitimacy. In Berlin and Nice, the trials have been conducted where volunteers have signed up to be “members of the public” to test the facial recognition technology.

By contrast, in the Met police trial in Romford, he saw one young man targeted after being seen to avoid the facial recognition cameras which were signposted as in operation. “He was stopped and searched and had a small quantity of cannabis on him and was arrested,” he said.

He may have acted suspiciously by trying to avoid the camera but he was not a suspect on any “watch list,” nor had he consented to take part in the “trial.”

“For me, one of the big issues is around the definition of ‘wanted’,” said Professor Fussey. “There is ‘wanted by the courts’ where someone has absconded and there is judicial oversight. Then there is ‘wanted by the police’ which is wanted for questioning or wanted on the basis of suspicion.”

South Wales police was careful to prepare four watchlists: red – those who posed a risk to public safety, amber – offenders with previous convictions, green – those whose presence did not pose any risk, blue – police officers’ faces to test the system.

However, human rights campaigners cite as an example police databases that hold the images of 21 million innocent people who have been arrested or in custody but never convicted.

It has been ruled such databases are illegal but so great is the task of processing and deleting them that progress on doing so has stalled.

So concerned is the Commons science and technology committee that in its recent report on face recognition it called for a moratorium on its use until rules on its deployment are agreed.

Professor Wiles sums it up:

“There are some uses that are not in the public interest. What that raises is who should make that decision about those uses. The one thing I am clear about is that the people who want to use facial recognition technology should not be the people who make that decision. It ought to be decided by a body that represents the public interest and the most obvious one is Parliament. There should be governance backed by legislation. Parliament should decide, yes, this is in the public interest provided these conditions are met. We have done that with DNA. Parliament said it’s in the public interest that the police can derive profiles of individuals from DNA but it’s not in the public interest that police could keep the samples. You can tell a lot more about a person from a sample. It’s important because if you get it wrong, the police will lose public trust and in Britain, we have a historic tradition of policing by consent.”

Ms Denham, the Information Commissioner, is to publish the results of her investigation into facial recognition, which she says should only be deployed where there is “demonstrable evidence” that it is “necessary, proportionate and effective.”

She has demanded police and other organisations using facial technology must ensure safeguards are in place including assessments of how it will affect people before each deployment and a clear public policy on why it is being used.

“There remain significant privacy and data protection issues that must be addressed and I remain deeply concerned about the rollout of this technology,” she says.

We are hurtling towards a surveillance state’: the rise of facial recognition technology

It can pick out shoplifters, international criminals and lost children in seconds. But as the cameras proliferate, who’s watching the watchers?

Cameras making up the image of a face against an orange background
 ‘If you’ve got something to be worried about, you should probably be worried.’ Photograph: Lol Keegan/The Guardian. Cameras supplied by

Gordon’s wine bar is reached through a discreet side-door, a few paces from the slipstream of London theatregoers and suited professionals powering towards their evening train. A steep staircase plunges visitors into a dimly lit cavern, lined with dusty champagne bottles and faded newspaper clippings, which appears to have had only minor refurbishment since it opened in 1890. “If Miss Havisham was in the licensing trade,” an Evening Standard review once suggested, “this could have been the result.”

The bar’s Dickensian gloom is a selling point for people embarking on affairs, and actors or politicians wanting a quiet drink – but also for pickpockets. When Simon Gordon took over the family business in the early 2000s, he would spend hours scrutinising the faces of the people who haunted his CCTV footage.

“There was one guy who I almost felt I knew,” he says. “He used to come down here the whole time and steal.” The man vanished for a six-month stretch, but then reappeared, chubbier, apparently after a stint in jail. When two of Gordon’s friends visited the bar for lunch and both had their wallets pinched in his presence, he decided to take matters into his own hands. “The police did nothing about it,” he says. “It really annoyed me.”

Gordon is in his early 60s, with sandy hair and a glowing tan that hints at regular visits to Italian vineyards. He makes an unlikely tech entrepreneur, but his frustration spurred him to launch Facewatch, a fast-track crime-reporting platform that allows clients (shops, hotels, casinos) to upload an incident report and CCTV clips to the police. Two years ago, when facial recognition technology was becoming widely available, the business pivoted from simply reporting into active crime deterrence. Nick Fisher, a former retail executive, was appointed Facewatch CEO; Gordon is its chairman.

Gordon installed a £3,000 camera system at the entrance to the bar and, using off-the-shelf software to carry out facial recognition analysis, began collating a private watchlist of people he had observed stealing, being aggressive or causing damage. Almost overnight, the pickpockets vanished, possibly put off by a warning at the entrance that the cameras are in use.

The company has since rolled out the service to at least 15 “household name retailers”, which can upload photographs of people suspected of shoplifting, or other crimes, to a centralised rogues’ gallery in the cloud. Facewatch provides subscribers with a high-resolution camera that can be mounted at the entrance to their premises, capturing the faces of everyone who walks in. These images are sent to a computer, which extracts biometric information and compares it to faces in the database. If there’s a close match, the shop or bar manager receives a ping on their mobile phone, allowing them to monitor the target or ask them to leave; otherwise, the biometric data is discarded. It’s a process that takes seconds.

Facewatch HQ is around the corner from Gordon’s, brightly lit and furnished like a tech company. Fisher invites me to approach a fisheye CCTV camera mounted at face height on the office wall; he reassures me that I won’t be entered on to the watchlist. The camera captures a thumbnail photo of my face, which is beamed to an “edge box” (a sophisticated computer) and converted into a string of numbers. My biometric data is then compared with that of the faces on the watchlist. I am not a match: “It has no history of you,” Fisher explains. However, when he walks in front of the camera, his phone pings almost instantly, as his face is matched to a seven-year-old photo that he has saved in a test watchlist.

“If you’re not a subject of interest, we don’t store any images,” Fisher says. “The argument that you walk in front of a facial recognition camera, and it gets stored and you get tracked is just.” He pauses. “It depends who’s using it.”

While researching theft prevention, Fisher consulted a career criminal from Leeds who told him that, for people in his line of work, “the holy grail is, don’t get recognised”. This, he says, makes Facewatch the ultimate deterrent. He tells me he has signed a deal with a major UK supermarket chain (he won’t reveal which) and is set to roll out the system across their stores this autumn. On a conservative estimate, Fisher says, Facewatch will have 5,000 cameras across the UK by 2022.

The company also has a contract with the Brazilian police, who have used the platform in Rio de Janeiro.

“We caught the number two on Interpol’s most-wanted South America list, a drug baron,”

says Fisher, who adds the system also led to the capture of a male murderer who had been on the run for several years, spotted dressed as a woman at the Rio carnival. I ask him whether people are right to be concerned about the potential of facial recognition to erode personal privacy.

“My view is that, if you’ve got something to be worried about, you should probably be worried,” he says. “If it’s used proportionately and responsibly, it’s probably one of the safest technologies today.

Unsurprisingly, not everyone sees things this way. In the past year, as the use of facial recognition technology by police and private companies has increased, the debate has intensified over the threat it could pose to personal privacy and marginalised groups.

The cameras have been tested by the Metropolitan police at Notting Hill carnivala Remembrance Sunday commemoration, and at the Westfield shopping centre in Stratford, east London. This summer, the London mayor, Sadiq Khan, wrote to the owners of a private development in King’s Cross, demanding more information after it emerged that facial recognition had been deployed there for unknown purposes.

In May, Ed Bridges, a public affairs manager at Cardiff University, launched a landmark legal case against South Wales police. He had noticed facial recognition cameras in use while Christmas shopping in Cardiff city centre in 2018. Bridges was troubled by the intrusion. “It was only when I got close enough to the van to read the words ‘facial recognition technology’ that I realised what it was, by which time I would’ve already had my data captured and processed,” he says. When he noticed the cameras again a few months later, at a peaceful protest in Cardiff against the arms trade, he was even more concerned: it felt like an infringement of privacy, designed to deter people from protesting. South Wales police have been using the technology since 2017, often at major sporting and music events, to spot people suspected of crimes, and other “persons of interest”. Their most recent deployment, in September, was at the Elvis Festival in Porthcawl.

“I didn’t wake up one morning and think, ‘I want to take my local police to court’,” Bridges says. “The objection I had was over the way they were using the technology. The police in this country police by consent. This undermines trust in them.”

Nick Fisher, CEO of Facewatch, a UK facial-recognition firm, in Gordon’s Wine Bar, London
 Nick Fisher, CEO of Facewatch, a UK facial-recognition firm that started life as a way to track pickpockets in a London wine bar. Photograph: Karen Robinson/The Guardian

During a three-day hearing, lawyers for Bridge, supported by the human rights group Liberty, alleged the surveillance operation breached data protection and equality laws. But last month, Cardiff’s high court ruled that the trial, backed by £2m from the Home Office, had been lawful. Bridges is appealing, but South Wales police are pushing forward with a new trial of a facial recognition app on officers’ mobile phones. The force says it will enable officers to confirm the identity of a suspect “almost instantaneously, even if that suspect provides false or misleading details, thus securing their quick arrest”.

The Metropolitan police have also been the subject of a judicial review by the privacy group Big Brother Watch and the Green peer Jenny Jones, who discovered that her own picture was held on a police database of “domestic extremists”.

In contrast with DNA and fingerprint data, which normally have to be destroyed within a certain time period if individuals are arrested or charged but not convicted, there are no specific rules in the UK on the retention of facial images. The Police National Database has snowballed to contain about 20m faces, of which a large proportion have never been charged or convicted of an offence. Unlike DNA and fingerprints, this data can also be acquired without a person’s knowledge or consent.

“I think there are really big legal questions,” says Silkie Carlo, director of Big Brother Watch. “The notion of doing biometric identity checks on millions of people to identify a handful of suspects is completely unprecedented. There is no legal basis to do that. It takes us hurtling down the road towards a much more expansive surveillance state.”

Some countries have embraced the potential of facial recognition. In China, which has about 200m surveillance cameras, it has become a major element of the Xue Liang (Sharp Eyes) programme, which ranks the trustworthiness of citizens and penalises or credits them accordingly. Cameras and checkpoints have been rolled out most intensively in the north-western Xinjiang province, where the Uighur people, a Muslim and minority ethnic group, account for nearly half the population. Face scanners at the entrances of shopping malls, mosques and at traffic crossings allow the government to cross-reference with photos on ID cards to track and control the movement of citizens and their access to phone and bank services.

At the other end of the spectrum, San Francisco became the first major US city to ban police and other agencies from using the technology in May this year, with supervisor Aaron Peskin saying: “We can have good policing without being a police state.”

Meanwhile, the UK government has faced harsh criticism from its own biometrics commissioner, Prof Paul Wiles, who said the technology is being rolled out in a “chaotic” fashion in the absence of any clear laws. Brexit has dominated the political agenda for the past three years; while politicians have looked the other way, more and more cameras are being allowed to look at us.

Facial recognition is not a new crime-fighting tool. In 1998, a system called FaceIt, comprising a handful of CCTV cameras linked to a computer, was rolled out to great fanfare by police in the east London borough of Newham. At one stage, it was credited with a 40% drop in crime. But these early systems only worked reliably in the lab. In 2002, a Guardian reporter tried in vain to get spotted by FaceIt after police agreed to add him to their watchlist. He compared the system to a fake burglar alarm on the front of a house: it cuts crime because people believe it works, not because it does.

However, in the past three years, the performance of facial recognition has stepped up dramatically. Independent tests by the US National Institute of Standards and Technology (Nist) found the failure rate for finding a target picture in a database of 12m faces had dropped from 5% in 2010 to 0.1% this year.

The rapid acceleration is thanks, in part, to the goldmine of face images that have been uploaded to Instagram, Facebook, LinkedIn and captioned news articles in the past decade. At one time, scientists would create bespoke databases by laboriously photographing hundreds of volunteers at different angles, in different lighting conditions. By 2016, Microsoft had published a dataset, MS Celeb, with 10m face images of 100,000 people harvested from search engines – they included celebrities, broadcasters, business people and anyone with multiple tagged pictures that had been uploaded under a Creative Commons licence, allowing them to be used for research. The dataset was quietly deleted in June, after it emerged that it may have aided the development of software used by the Chinese state to control its Uighur population.

In parallel, hardware companies have developed a new generation of powerful processing chips, called Graphics Processing Units (GPUs), uniquely adapted to crunch through a colossal number of calculations every second. The combination of big data and GPUs paved the way for an entirely new approach to facial recognition, called deep learning, which is powering a wider AI revolution.

“The performance is just incredible,” says Maja Pantic, research director at Samsung AI Centre, Cambridge, and a pioneer in computer vision. “Deep [learning] solved some of the long-standing problems in object recognition, including face recognition.”

Recognising faces is something like a game of snap – only with millions of cards in play rather than the standard deck of 52. As a human, that skill feels intuitive, but it turns out that our brains perform this task in a surprisingly abstract and mathematical way, which computers are only now learning to emulate. The crux of the problem is this: if you’re only allowed to make a limited number of measurements of a face – 100, say – what do you choose to measure? Which facial landmarks differ most between people, and therefore give you the best shot at distinguishing faces?

A deep-learning program (sometimes referred to more ominously as an “agent”) solves this problem through trial and error. The first step is to give it a training data set, comprising pairs of faces that it tries to match. The program starts out by making random measurements (for example, the distance from ear to ear); its guesses will initially be little better than chance. But at each attempt, it gets feedback on whether it was right or wrong, meaning that over millions of iterations it figures out which facial measurements are the most useful. Once a program has worked out how to distil faces into a string of numbers, the algorithm is packaged up as software that can be sent out into the world, to look at faces it has never seen before.

The performance of facial recognition software varies significantly, but the most effective algorithms available, such as Microsoft’s, or NEC’s NeoFace, very rarely fail to match faces using a high-quality photograph. There is far less information, though, about the performance of these algorithms using images from CCTV cameras, which don’t always give a clear view.

Recent trials reveal some of technology’s real-world shortcomings. When South Wales police tried out their NeoFace system for 55 hours, 2,900 potential matches were flagged, of which 2,755 were false positives and just 18 led to arrests (the number charged was not disclosed). One woman on the watchlist was “spotted” 10 times – none of the sightings turned out to be of her. This led to claims that the software is woefully inaccurate; in fact, police had set the threshold for a match at 60%, meaning that faces do not have to be rated as that similar to be flagged up. This minimises the chance of a person of interest slipping through the net, but also makes a lot of false positives inevitable.

In general, Pantic says, the public overestimates the capabilities of facial recognition. In the absence of concrete details about the purpose of the surveillance in London’s King’s Cross this summer, newspapers speculated that the cameras could be tracking shoppers and storing their biometric data. Pantic dismisses this suggestion as “ridiculous”. Her own team has developed, as far as she is aware, the world’s leading algorithm for learning new faces, and it can only store the information from about 50 faces before it slows down and stops working. “It’s huge work,” she says. “People don’t understand how the technology works, and start spreading fear for no reason.”

This week, the Met police revealed that seven images of suspects and missing people had been supplied to the King’s Cross estate “to assist in the prevention of crime”, after earlier denying any involvement. Writing to the London Assembly, the deputy London mayor, Sophie Linden, said she “wanted to pass on the [Metropolitan police service’s] apology” for failing to previously disclose that the scheme existed, and announced that similar local image-sharing agreements were now banned. The police did not disclose whether any related arrests took place.

Like many of those working at the sharp end of AI, Pantic believes the controversy is “super overblown”. After all, she suggests, how seriously can we take people’s concerns when they willingly upload millions of pictures to Facebook and allow their mobile phone to track their location? “The real problem is the phones,” she says – a surprising statement from the head of Samsung’s AI lab. “You are constantly pushed to have location services on. [Tech companies] know where you are, who you are with, what you ate, what you spent, wherever you are on the Earth.”

Concerns have been raised that facial recognition has a diversity problem, after widely cited research by MIT and Stanford University found that software supplied by three companies misassigned gender in 21% to 35% of cases for darker-skinned women, compared with just 1% for light-skinned men. However, based on the top 20 algorithms, Nist found that there is an average difference of just 0.3% in accuracy between performance for men, women, light- and dark-skinned faces. Even so, says Carlo of Big Brother Watch, the technology’s impact could still be discriminatory because of where it is deployed and whose biometric data ends up on databases. It’s troubling, she says, that for two years, Notting Hill carnival, the country’s largest celebration of Caribbean and black British culture, was seen as an “acceptable testing ground” for the technology.

Maja Pantic, research director of Samsung Al Research Centre, photographed at Imperial College, London
 ‘The real problem is phones’: Maja Pantic, research director at Samsung’s AI Centre. Photograph: Karen Robinson/The Guardian

I ask Fisher about the risk of racial profiling: the charge that some groups may be more likely to fall under suspicion, say, when a shop owner is faced with ambiguous security footage. He dismisses the concern. Facewatch clients are required to record the justification for their decision to upload a picture on to the watchlist and, in a worst-case scenario, he argues, a blameless individual might be approached by a shopkeeper, not thrown into jail.

“You’re talking about human prejudices, you can’t blame the technology for that,” he says.

After our interview, I email several times to ask for a demographic breakdown of the people on the watchlist, which Fisher had offered to provide; Facewatch declines.

Bhuwan Ribhu grew up in Delhi, in a small apartment with his parents, his sister Asmita, and many children who had been rescued from slavery and exploitation. Like Gordon, Ribhu followed his parents into the family business – in his case, tracking down India’s missing children, who have been enticed, forcibly taken or sold by their parents to traffickers, and end up working in illegal factories, quarries, farms and brothels. His father is the Nobel Peace laureate Kailash Satyarthi, who founded the campaign Bachpan Bachao Andolan (Save Childhood Movement) in 1980, after realising that he could not accommodate all of the children being rescued in the family home.

The scale of the challenge is almost incomprehensible: 63,407 child kidnappings were reported to Indian police in 2016, according to the National Crime Records Bureau. Many children later resurface, but the sheer numbers involved mean it can take months or years to reunite them with their families. “About 300,000 children have gone missing over the last five or six years, and 100,000 children are housed in various childcare institutions,” says Ribhu. “For many of those, there is a parent out there looking for their child. But it is impossible to manually go through them all.”

He describes the case of Sonu, a boy from India’s rural Bihar region, 1,000km from Delhi. When Sonu was 13, his parents entrusted him to a factory owner who promised him a better life and money. But they quickly lost track of their son’s whereabouts and began to fear for his safety. Eventually they contacted Bachpan Bachao Andolan for help. Sonu was tracked down after about two years, hundreds of miles from home. “We found the child after sending out his photo to about 1,700 childcare institutions across India,” Ribhu says. “One of them called us back and said they might have the child. People went and physically verified it. We were looking for one child in a country of 1.3 billion.”

Ribhu had read a newspaper article about the use of facial recognition to identify terrorists at airports and realised it could help. India has created two centralised databases in recent years: one containing photos of missing children, and the other containing photos of children housed in childcare institutions. In April last year, a trial was launched to see whether facial recognition software could be used to match the identities of missing and found children in the Delhi region. The trial was quickly hailed a success, with international news reports suggesting that “nearly 3,000 children” had been identified within four days. This was an exaggeration: the 3,000 figure refers to potential matches flagged by the software, not verified identifications, and it proves difficult to find out how many children have been returned to parents. (The Ministry of Women and Child Development did not respond to questions.) But Ribhu says that, since being rolled out nationally in April, there have been 10,561 possible matches and the charity has “unofficial knowledge” of more than 800 of these having been verified. “It has already started making a difference,” he says. “For the parents whose child has been returned because of these efforts, for the parents whose child has not gone missing because the traffickers are in jail. We are using all the technological solutions available.”

Watching footage of Sonu being reunited with his parents in a recent documentary, The Price Of Free, it is hard to argue against the deployment of a technology that could have ended his ordeal more quickly. Nonetheless, some privacy activists say such successes are used to distract from a more open-ended surveillance agenda. In July, India’s Home Ministry put out a tender for a new Automated Facial Recognition System (AFRS) to help use real-time CCTV footage to identify missing children – but also criminals and others, by comparing the footage with a “watchlist” curated from police databases or other sources.

Real-time facial recognition, if combined with the world’s largest biometric database (known as Aadhaar), could create the “perfect Orwellian state”, according to Vidushi Marda, a legal researcher at the human rights campaign group Article 19. About 90% of the Indian population are enrolled in Aadhaar, which allocates people a 12-digit ID number to access government services, and requires the submission of a photograph, fingerprints and iris scans. Police do not currently have access to Aadhaar records, but some fear that this could change.

“If you say we’re finding missing children with a technology, it’s very difficult for anyone to say, ‘Don’t do it’,” Marda says. “But I think just rolling it out now is more dangerous than good.”

Debates about civil liberties are often dictated by instinct: ultimately, how much do you trust law enforcement and private companies to do the right thing? When searching for common ground, I notice that both sides frequently reference China as an undesirable endpoint. Fisher thinks that the recent disquiet about facial recognition stems from the paranoia people feel after reading about its deployment there. “They’ve created digital prisons using facial recognition technology. You can’t use your credit card, you can’t get a taxi, you can’t get a bus, your mobile phone stops working,” he says. “But that’s China. We’re not China.”

Groups such as Liberty and Big Brother Watch say the opposite: since facial recognition, by definition, requires every face in a crowd to be scanned to identify a single suspect, it will turn any country that adopts it into a police state. “China has made a strategic choice that these technologies will absolutely intrude on people’s liberty,” says biometrics commissioner Paul Wiles. “The decisions we make will decide the future of our social and political world.”

For now, it seems that the question of whether facial recognition will make us safer, or represents a new kind of unsafe, is being left largely to chance. “You can’t leave [this question] to people who want to use the technology,” Wiles says. “It shouldn’t be the owners of the space around King’s Cross, it shouldn’t be Facewatch, it shouldn’t be the police or ministers alone – it should be parliament.”

After leaving the Facewatch office, I walk along the terrace of Gordon’s, where a couple of lunchtime customers are enjoying a bottle of red in the sunshine, and past the fisheye lens at the entrance to the bar, which I now know is beaming my face to the computer cloud. I think back to a winky promise I’ve read on the Gordon’s website: “Make your way to the cellar to your rickety candlelit table – anonymity is guaranteed!”

Out in the wider world, anonymity is no longer guaranteed. Facial recognition gives police and companies the means of identifying and tracking people of interest, while others are free to go about their business. The real question is: who gets that privilege?