AI Takes Over the Cameras Watching Our Cities to Prevent Crime — and Makes Orwellian Surveillance Possible | AI: The Key Is People ③

To be safe, you have to give up your privacy. Right or wrong, there is no going back.

In 1962, nothing was ever stolen in New York. Apartment doors didn't even have locks.

But every apartment had a camera hidden in it, and the picture went to SS headquarters. In a dim room, SS soldiers in black uniforms watched a wall of small screens, recording every resident's every move.

This is the new world imagined by the American series The Man in the High Castle, in which the Allies lost the Second World War and New York became the capital of a Nazi America. Life there is safer for people, and more dangerous.

Whatever kind of film you like, you have probably seen this scene. Depending on what the plot needs, the watcher in front of the monitor is either fooled easily or catches someone in the act.

If the plan of the graphics-chip giant Nvidia comes true, though, that cliché will get rarer and rarer.

Last month the chip designer Nvidia announced an artificial-intelligence platform called Metropolis. The name sounds thoroughly cyberpunk, and the product behind it offers an equally high-tech vision: the surveillance cameras of an entire city would stop being separate closed-circuit systems and instead all run on one platform, with every video stream pooled in cloud storage and analyzed by artificial intelligence.

Once trained, the AI would take the place of human officers in picking out every incident happening in a city — traffic accidents, violent crime, fires. Under Nvidia's plan, all of it would be routed to the city's management departments to speed up the response of the police and the fire brigade.

Nvidia has assembled a whole supply chain for this vision. It is working with traditional surveillance-camera makers such as Dahua Technology and Hikvision so that their cameras use Nvidia chips; the AI computing that follows is done by Nvidia's Jetson chip; and the platform's cloud computing is accelerated by Nvidia's DGX-1 supercomputing modules.

If governments really pay for this platform and put it to work, what it can do will probably go well beyond a traditional surveillance camera. The AI can not only recognize a face; it can also identify a person by how they move and what they are carrying.

Say you are wearing a mask and a hat. Future AI surveillance could still identify you by your walk. The system would know that you left home, threw out a bag of rubbish and then went into the subway. If you were a terrorism suspect, that bag of rubbish might be flagged as a possible bomb and set the police in motion.

In other words, a surveillance system would no longer just see a person's face. It would be able to watch every move a person makes.

Part of what Nvidia has planned is in fact already here. Today's smart surveillance systems still cannot work out what a person is doing, but keeping track of every person and every car and photographing them is mature technology.

A 2014 promotional video from the Chinese smart-surveillance company Deepglint showed a system that could identify every car and every person on the road and photograph and file them automatically.

The Groundwork Is Already Here

The usual form of urban video surveillance today is closed-circuit television, or CCTV. In such a system the pictures a camera takes are never broadcast publicly; they travel and are stored inside a closed system — the room full of screens you see in films.

In any big city in the world, surveillance cameras are now everywhere. Take Beijing, a special case: cameras belonging to at least three government departments are working there all the time.

The most common kind is public-order surveillance — the small cameras on every street corner that look like streetlights. They sit on pan-tilt mounts and can swivel, they are generally run by the police, and their job is to watch over everyday public order.

The bullet-shaped cameras above traffic lights are mainly there to photograph cars and pedestrians breaking the rules. On motorways and expressways, some of the bullet cameras mounted on overhead gantries also measure speed. These belong to the traffic police.

Then there is a third kind, mounted very high up and shaped like a lamppost. Some of these belong to traffic-management departments and mostly watch how the traffic is flowing.

Over the past few years, "smart city" projects and the like have upgraded these cameras so that they can capture faces and number plates clearly. Whether you are on foot, in a car or on a bicycle, some camera or other will get a picture of you.

It all started in London, the first city in the world to be covered by CCTV. Its first fixed surveillance system went into service in 1968, installed near the American embassy on Grosvenor Square to watch students protesting the Vietnam War.

After that, surveillance systems spread. In 1993, closed-circuit footage helped solve a murder: a boy named James Bulger disappeared from a shopping centre in northwest England and was later killed. The footage showed two older boys leading him away. On the strength of it, the two were charged with murder, and the last images of James Bulger caught by the cameras shocked the public. CCTV won wide support as a result.

By the end of the 1990s, on the theory that surveillance prevents crime, more than half of the British Home Office's crime-prevention budget was going into building out surveillance cameras.

Measured by density of coverage, Britain has become the country with the most cameras per person in the world: 4 million to 5.9 million cameras on its streets and in its public spaces, one camera for every 11 Britons, and roughly 300 CCTV shots of each person's face every day.

Figures from 2015 showed more than 420,000 surveillance cameras in London, second in the world among cities. First was Beijing: in 2015 state media reported, approvingly, that urban Beijing had reached 100% camera coverage, with more than 470,000 cameras.

The most ironic part is the name of the programme that produced all those cameras: Skynet. The Skynet Project was a national video-surveillance push, promoted before 2008 by the Ministry of Public Security, the Political and Legal Affairs Commission and the then Ministry of Information Industry, meant to computerize public-order management and population records. Whoever chose the name was probably not thinking of Skynet, the superintelligent AI that enslaves humanity in The Terminator. But in Chinese cities and villages the programme has, in effect, achieved coverage of the entire population.

The official Beijing figure of 470,000 cameras is probably nowhere near the real one, since these projects do not have to be disclosed to the public. The number of cameras actually working can only be that or higher.

On crime-and-law programmes on television, you often see officers deducing a suspect's identity from a blurry piece of video. In reality, government-funded surveillance cameras stopped being that blurry long ago. These cameras now commonly shoot at 2 to 3 megapixels, and some models reach 28 megapixels — getting a clear picture of a face inside a moving car is easy.

Better technology has made enforcement more exact. In early 2017 the traffic police announced that, from the second half of the year, when a camera catches a car breaking the rules, the points can no longer be docked from just anyone's licence, as they were before; the penalty has to go against the licence of whoever was driving at the time. That is because traffic cameras have been upgraded to capture the driver's face along with the number plate.

In Britain, the argument over surveillance in public places has run for years. In 1994 the home secretary of the day tried to reassure the public by calling surveillance cameras "benevolent eyes in the sky" and saying they did no harm to law-abiding citizens.

The objections never stopped. In 2007 Richard Thomas, then Britain's information commissioner, warned that "excessive use of CCTV and other information-gathering systems creates a climate of suspicion in society," and the old theory that surveillance deters crime came under doubt.

In 2016 the European Court of Justice ruled that Britain's Data Retention and Investigatory Powers Act 2014 was unlawful, holding that large-scale, indiscriminate video surveillance of the public was illegal and that such surveillance required a court's permission in a specific case.

But the law that was struck down was due to lapse at the end of 2016 anyway, replaced by the Investigatory Powers Act. The new law gives nearly 50 British government departments the right to look at surveillance data. The governing Labour Party argued that the move would help fight terrorism and crime.

With the technology of the moment, though, the system was as poor at watching ordinary citizens as it was at watching suspects.

Today's City Surveillance Systems Still Run Mostly on Human Eyes

In 2011 riots broke out across London. Crowds of rioters took to the streets and faced down the police. After the police used force to put down the riots and clear the streets, they still had to find the people who had destroyed property, burned cars and looted shops. In the end London police identified 5,000 suspects from surveillance footage.

Five thousand sounds good, but it took a full five months, because there were 200,000 hours of footage.

The special police unit that does this work is called the "super-recognizers." On top of patrolling and making arrests, their main job is to look at enormous quantities of low-resolution faces from surveillance cameras and pick out repeat offenders.

Officer Gary Collins, for one, picked out a young man who had firebombed a car.

The video was shot at night and was low-resolution to begin with. The man wore a wool hat, the light caught only his profile, and one side of his face was in shadow. Collins recognized him anyway: a robber he had arrested before. He went back to the old arrest record, found the man's name and brought him in.

Collins and his squad have an unusual gift for faces. On the Cambridge Face Memory Test they can recognize 95% of the faces they have seen, while an ordinary person scores only 20% (if you don't believe it, take the test yourself).

There are too few people who never forget a face, and far too much video to get through.

In the British police the "super-recognizer" team has only six members, and they lead a trained group of 150 mobile officers to handle the hardest footage.

It is like Will Smith's early film Enemy of the State, where the National Security Agency's men already have all sorts of implausible technology — satellites on demand, access to city systems — and still need a room full of people staring at screens to track the hero down.

Which is why, even in Britain, with the tightest surveillance system anywhere, the footage can only deter and help with the investigation afterwards. Neither the police nor the shopkeepers who install the cameras have the people to spot a problem on screen as it happens and react.

Even the suspicious behaviour before Britain's recent string of terrorist attacks passed the surveillance system by. The threat the system poses to people's privacy exists only as a theoretical possibility.

For the Past Few Years, Tech Companies Have Been Trying to Use AI to Make Cameras Smart

Lately that has begun to change. "Closed"-circuit television is no longer closed.

In 2014, after the Boston Marathon bombing, Boston police started replacing the city's surveillance cameras. They used video-analysis technology from BRS Labs, then a Canadian company, which lets the machine learn for itself what kind of picture looks suspicious.

These new-generation "closed"-circuit systems work like this: first the system watches a setting with both people and cars in it and produces a result; a human monitor then confirms whether the "suspicious activity" it found is really suspicious and sends the answer back to the algorithm. After a period of training, the algorithm analyzes the footage it captures directly, as it comes in.

With technology like that of BRS Labs, a company does not need to replace its hardware. As long as the cameras can connect to a network, they can run the algorithm.

BRS raised five rounds of funding, took in $23 million, was bought twice, and is now one of the solutions sold by the deep-learning tools company OmniAI.

Its product was later used for traffic and public-space video surveillance in several American cities — on San Francisco buses, for example.

What the traditional surveillance companies are good at is less technology than winning government contracts.

Getting surveillance systems to exchange data across regions and across different kinds of information has gone slowly. The contractors' abilities are usually limited, and cooperation between government departments is slow too. Under an earlier plan by China's Ministry of Public Security, for instance, it was not until 2015 that footage could be pulled from another province.

The arrival of a technology company like Nvidia could change that. The self-driving system it built with Tesla is already in the new Model S, and its main business, graphics-processing chips, is also the hardware best suited to image recognition today.

Its founder, Jensen Huang, has spent a year saying the company is turning into an artificial-intelligence company.

Traditional CCTV makers have joined this smart-surveillance rush over the past two years as well.

For a company that sells cameras to governments, working with Nvidia means the hardest parts — data centres, algorithms, management software — can be handed off, while it keeps hold of what it does best: government relations.

Nvidia is not the only giant interested in artificial intelligence. Intel spent more than $10 billion last year buying MobilEye, the company that helped Tesla build its first-generation self-driving system, and it is exploring other areas of AI as well. Since Google's open-source AI platform TensorFlow went live, startups have been able to call directly on Google's speech-recognition and translation computing.

Beyond hardware and algorithms, face recognition has come a long way — not only the algorithms but the databases, which are bigger than they have ever been. Facebook has spent years encouraging users to tag friends in photos, and Google and Apple both recognize faces automatically in their own photo apps; they simply don't make the results public.

And the American government, which runs no social network of its own, already has half the adult population in its database.

So... Do We Have to Accept This New Reality?

The terrorist attacks and social troubles of recent years have given governments and companies a reason to make the public "afraid" enough to let smart surveillance equipment — with video analysis and face recognition — into everyday life.

There is something to the argument. Smart cameras can analyze suspicious behaviour in public places as it happens, recognize criminals as it happens, and tell the police while a crime is being committed or even before. Such a system would let the police react earlier, and it would frighten people off more effectively.

But a system that can watch everyone, and will eventually recognize everyone, may bring not only safety but the total surveillance of 1984.

"The innocent have nothing to fear" is a line the supporters of surveillance use often.

Glenn Greenwald, the Guardian reporter who first broke the Snowden story, has a favourite reply: "Would you write down all your account passwords and email them to me?"

Almost nobody would. Yet almost everybody is handing more and more information to companies and to governments. It used to be fingerprints and digital records. With the help of artificial intelligence, it is now every minute and every second after you leave the house.

As for the privacy protection you can expect right now, here is the rule in Britain, the country with the most cameras: you can ask for a copy of surveillance footage that shows your face, but you cannot have the original deleted.

Some large technology companies are still restraining themselves to keep their users' trust. Google's and IBM's systems can pick many people out of an image, for instance, but they recognize without returning the result — celebrities excepted. Not every company has made a similar promise, and a promise is only a promise.

"Giving up privacy does not necessarily bring safety. Better security does not necessarily require losing privacy," said Jennifer Granick, then director of the Center for Internet and Society at Stanford Law School.

In her view, the problem starts with governments and companies making the public "pick one" between privacy and security, or between privacy and service. What is urgent is a privacy law — rules setting out how companies and governments may use it.

Granick said that in May 2006. The decade since has brought little progress. The right to erase your personal information from the digital world exists only in Europe (the right-to-be-forgotten lawsuit of 2015).

Somebody else's words did come true.

In 2010 Mark Zuckerberg — the man who made social networking universal and got people to give up their digital privacy willingly — said that having no privacy was the "new social norm" of the future.

Three years later, Zuckerberg, worth $100 billion after his company's listing, spent $30 million buying the four houses next to his own to protect his privacy in real life.

In the new world, privacy really isn't something an ordinary person can afford.