Surveillance Of People Is More "Pervasive And Normalised" Than Previously Thought, Endangering Our Privacy

Research into computer vision development – a field of artificial intelligence (AI) where computers learn to “see” and interpret information and images – has significantly contributed to pervasive surveillance of people and their habits. In addition, it seems this rise in surveillance has coincided with a rise in obfuscating language that either normalizes or attempts to hide the existence of the surveillance.

When it comes to gathering information, surveillance agencies have a few options. Firstly, they can collect the data directly from the infrastructure carrying internet traffic. This is called “upstream surveillance” and tends to involve intercepting information directly from things like routers, switches, or undersea cables. This allows them to access a huge volume of data from emails, search engine queries, metadata, and browser activity in real time as it flows through the internet.

Then there’s “downstream surveillance”, whereby authorities can gain data directly from service providers like Google, Facebook, or Yahoo. So rather than intercepting information through a checkpoint, the authority accesses stores and real-time communications from these service providers, usually with their cooperation after receiving a warrant (in theory).

Needless to say, these are both controversial practices, but new research has discovered a more worrying increase in the latter form of surveillance from an unsuspecting source: the field of computer vision (CV).

At its heart, CV involves analyzing digital images and videos for meaningful information in order to identify, interpret, and understand the world in a similar way to humans. The process can involve training the AI to recognize faces or objects, tracking motion, detecting anomalies, or potentially interpreting emotions. 

Originally tied to military and carceral surveillance, CV was initially intended to identify targets and gather intelligence for war, law enforcement, and immigration controls. However, over time, it has become the focus of a wider range of parties – including universities, funding agencies, governments, companies, and larger – who have shaped its trajectory through their own interests.

Generally speaking, only a small portion of CV research is thought to be harmful, but the latest study argues that surveillance is actually far more “pervasive and normalised” than we thought.

The research was conducted by scientists at Stanford University, Carnegie Mellon University, the University of Washington, and the Allen Institute for Artificial Intelligence. The team examined over 40,000 documents, CV papers, and downstream patents across four decades. They found that, across this time, there has been a five-fold increase in the number of CV papers linked to downstream surveillance patents.

“This work provides a detailed, systematic understanding of the field of computer vision research and presents a concrete empirical account that reveals the pathway from such research to surveillance and the extent of this surveillance,” Dr Abeba Birhane, director of the AI Accountability Lab in the ADAPT Research Ireland Centre at the School of Computer Science and Statistics in Trinity College Dublin, said in a statement.

Among the findings was that the field of CV has evolved linguistically. It has moved away from generic papers produced in the 1990s towards an elevated focus on analyzing semantic categories and human behavior in the 2010s. At the same time, surveillance has become increasingly hidden behind jargon and obfuscation, distracting those people at the center of the surveillance.

The researchers also identified the main institutions responsible for the most surveillance. These include: 1) Microsoft; 2) Carnegie Mellon University; 3) MIT; 4) the University of Illinois Urbana-Champaign; and 5) the Chinese University of Hong Kong. According to the assessment, the US is the number one nation for surveillance, followed by China and then the UK.

“Linguistically the field has increasingly adapted to obfuscate the existence and extent of surveillance. One such example is how the word ‘object’ has been normalised as an umbrella term which is often synonymous with ‘people’,” Birhane added.

“The most troublesome implications of this are that it is harder and harder to opt out, disconnect, or to ‘just be’, and that tech and applications that come from this surveillance are often used to access, monetise, coerce, and control individuals and communities at the margins of society.”

Ultimately, this research demonstrates how precarious our ideas of privacy and freedoms are.

“Due to pervasive and intensive data gathering and surveillance, our rights to privacy and related freedoms of movement, speech and expression are under significant threat,” Birhane stressed.

While it is easy to become fearful over this situation, Birhane and colleagues believe things are not set in stone. It is possible for resources like this large-scale, systematic study to aid regulators and policymakers in addressing these issues.

“We hope these findings will equip activists and grassroots communities with the empirical evidence they need to demand change, and to help transform systems and societies in a more rights-respecting direction,” Birhane explained.

“CV researchers could also adopt a more critical approach, exercise the right to conscientious objection, collectively protest and cancel surveillance projects, and change their focus to study ethical dimensions of the field, educate the public, or put forward informed advocacy.”

The paper is published in Nature.