Businesses and the government have spent years installing millions of surveillance cameras across the United States. Now, that technology is on the verge of getting a major upgrade, the American Civil Liberties Union warns in a new report.
Advancements in artificial intelligence could supercharge surveillance, allowing camera owners to identify “unusual” behavior, recognize actions like hugging or kissing, easily seek out embarrassing footage and estimate a person’s age or, possibly, even their disposition, the group argues.
“We face the prospect of an army of A.I. security guards being placed behind those lenses that are actually, in a meaningful way, monitoring us, making decisions about us, scrutinizing us,” said Jay Stanley, senior policy analyst at the A.C.L.U. and the author of the report, which was released on Thursday.
The United States is, by various estimates, home to tens of millions of surveillance cameras. While many of those devices have been around for years, it has been widely understood that it would be unfeasible, if not impossible, for each device to be constantly monitored and its footage carefully categorized and documented, Mr. Stanley notes in the report, titled “The Dawn of Robot Surveillance.” Even the Justice Department has said that watching such footage is “boring and mesmerizing,” and that attention fades after about 20 minutes.
But improvements to technology created to actively monitor such feeds, known by several names including “video analytics,” are poised to change that, ensuring that every second of footage can be analyzed.
“It honestly has both benefits and security consequences,” said Carl Vondrick, a professor of computer science at Columbia University, where he leads a group focused on computer vision and machine learning.
The ability to constantly analyze and learn from a video feed could help self-driving cars understand their surroundings, retail stores track their products and health professionals monitor their patients, he said. It can also be used to scrutinize the routines and actions of individuals on an enormous scale, the A.C.L.U. warns.
In the report, the organization imagined a handful of dystopian uses for the technology. In one, a politician requests footage of his enemies kissing in public, along with the identities of all involved. In another, a life insurance company offers rates based on how fast people run while exercising. And in another, a sheriff receives a daily list of people who appeared to be intoxicated in public, based on changes to their gait, speech or other patterns.
Analysts have valued the market for video analytics at as much as $3 billion, with the expectation that it will grow exponentially in the years to come. The important players include smaller businesses as well as household names such as Amazon, Cisco, Honeywell, IBM and Microsoft.
At a recent retail industry conference, IBM showed how its video analytics software could be used to count customers and estimate their ages and loyalty status, all in real time. The software could monitor the length of a line, identify a manager as he walked through a crowd, and flag people loitering outside the store.
Amazon’s Rekognition service, launched in 2016, can purportedly identify and track people, recognize celebrities and detect objects and read text. (The company drew criticism for pitching that service to law enforcement.) After employees protested, Google last year said it would not renew a contract with the Pentagon’s Project Maven, for which artificial intelligence is used to interpret video and images, potentially to improve the targeting of drone strikes.
What video analytics can do
Video analytics providers boast a range of capabilities, according to the A.C.L.U. report, including detection of objects dropped or left behind; analysis of a person’s direction, gait or movement; and even identification of attempts to enter a secure area by rushing in as another person enters or exits the space. Some companies say their services can discern demographic information or identify clothes and other objects, too.
Software is also being trained to identify a wide range of activities, such as using a phone, shaking hands, punching something, drinking beer and walking toward or away from an object. (Amazon claims that Rekognition can already identify some such actions, including “blowing out a candle” and “extinguishing fire.”)
One area of research that the A.C.L.U. described with particular concern is the movement to train software on “anomaly detection,” which can single out an individual for unusual, atypical or deviant behavior. Another is emotion recognition, which promises to discern a person’s mood, though there is little evidence that emotions are universal or can be determined by facial movements alone.
As the technology improves, users will be able to search videos by keyword, surfacing results for precise queries like “red car” or “man wearing hoodie,” a capability that already exists for images stored on Google Photos and Apple Photos.
The associated threat
The spread of such technology has a number of dangerous implications, the A.C.L.U. warns.
First, algorithms can be trained on biased data sets, leading to discriminatory results. The well-documented racial shortcomings of facial recognition technology, for example, have been linked to training data that skews heavily white and male.
Video analytics software is often trained on publicly available footage, such as YouTube videos, but there may be bias in the kinds of people who post them or in what such videos show.
“There are reasons to fear this technology when it works, and there are reasons to fear this technology when it doesn’t work,” Mr. Stanley said.
The use of video analytics may