17.9: Who Is Looking Now? Surveillance, Data, and Machine Vision

In the twenty-first century, cameras do not only serve human vision. They increasingly serve machines, corporations, governments, police departments, militaries, advertisers, border agencies, and algorithms. Photography now participates in systems that identify, track, classify, predict, and control. This marks a major shift in the history of the medium. Photographs are no longer made only to be looked at by people. Many images are now made for systems.

This does not mean that human viewers have disappeared. People still make, share, interpret, and respond to photographs. Yet a growing number of images are produced within environments where human looking is secondary. Surveillance cameras monitor streets, stores, schools, airports, workplaces, and private homes. Drones record landscapes, protests, border zones, and military targets from above. Satellites image the planet continuously. Facial recognition systems compare faces against databases. Machine-learning systems sort images into categories, identify objects, read license plates, track movement, and generate predictions. The camera has become part of a larger computational network.

This transformation changes what photography is. A traditional photograph might be understood as a visual representation: an image made for a viewer. A machine-readable image, by contrast, may function less as a picture than as data. It is processed, measured, compared, tagged, stored, and acted upon. The image may never be meaningfully “seen” by a person at all. It may be converted into coordinates, biometric templates, probability scores, metadata, behavioral patterns, or risk assessments. In this context, photography becomes operational.

The theorist Harun Farocki used the term “operational images” to describe images that do not primarily represent the world but participate in operations.¹ These are images that help machines or institutions do something: guide a missile, track a body, map a territory, sort a population, or control access. The concept is essential for understanding twenty-first-century photography. Many contemporary images are not made to be contemplated. They are made to function.

Surveillance cameras provide one of the most familiar examples. Closed-circuit cameras in public and private spaces record people continually, often without their active participation. These images may be watched live, stored for later review, or processed by automated systems. Their purpose is not memory in the personal sense, nor artistic expression, nor even necessarily evidence after the fact. Their purpose is behavioral management. The presence of the camera can shape how people move, gather, shop, protest, work, and inhabit public space.

Facial recognition intensifies this condition by turning the face into data. A portrait once promised likeness: an image through which a person might be remembered or recognized. Biometric systems transform facial likeness into measurable information. The face becomes something to be detected, extracted, matched, and scored. Simone Browne argues that contemporary surveillance technologies must be understood in relation to longer histories of racial formation and the policing of Black life, including slavery-era practices such as branding, runaway slave notices, and lantern laws.² Surveillance, in this sense, is not merely technological. It is historical.

This point is crucial because machine vision is often presented as neutral. Algorithms are frequently described as objective tools that simply detect patterns. Yet image-recognition systems are built from human decisions: what images are collected, how they are labeled, which categories are created, whose faces are included, whose faces are underrepresented, and what purposes the system is designed to serve. Machine vision does not escape culture. It absorbs it.

Artists and scholars have shown that these systems often reproduce social bias. Kate Crawford and Trevor Paglen’s project ImageNet Roulette exposed the troubling categories embedded in image datasets used to train artificial intelligence systems. The project allowed users to upload a photograph and receive a classification drawn from the ImageNet dataset, often revealing offensive, racialized, gendered, or socially loaded labels.³ The project demonstrated that machine vision is not simply a technical problem. It is a cultural and political problem because classification systems carry assumptions about what kinds of people and behaviors are recognizable, normal, suspicious, or valuable.

This is why the photograph as data must be treated carefully. A photograph may look like a familiar image on a screen, but within a computational system it may be broken into features that no longer resemble ordinary looking. Edges, distances, colors, ratios, textures, gestures, and spatial relationships become inputs. These inputs may then be used to make decisions: whether a phone unlocks, whether a person is admitted through a border checkpoint, whether a body is flagged as suspicious, whether an advertisement is targeted, whether a drone tracks a figure, or whether a police database generates a match.

The difference between images made for humans and images made for systems is therefore not only aesthetic. It is ethical and political. A human-facing image asks to be interpreted. A machine-facing image may trigger action. A photograph of a person walking down a street may be meaningless to a casual viewer, but to a surveillance system it can become location data, gait data, facial data, behavioral data, or evidence of association. The image becomes part of a profile.

Drones and satellites further expand this field of vision. From above, bodies, borders, military sites, forests, oceans, refugee camps, pipelines, and cities can be monitored at a distance. Remote vision can produce knowledge, but it can also produce power without accountability. The aerial image has a long history in warfare, mapping, policing, and colonial administration. In the twenty-first century, it has become cheaper, more automated, and more continuous. The sky itself has become part of photographic infrastructure.

Trevor Paglen’s work is central to this discussion because he repeatedly photographs systems designed to remain hidden. His projects have addressed classified military sites, surveillance satellites, undersea internet cables, artificial intelligence training datasets, and machine vision. Paglen’s studio describes his machine-vision works as involving images “designed to teach machines ‘how to see.’”⁴ This phrasing is important because it shifts attention away from what machines see and toward how they are trained to see. Vision is learned, even for machines.

Paglen’s Limit Telephotography series uses powerful lenses to photograph classified military installations from great distances. The resulting images are often hazy, abstract, and difficult to read. This difficulty is part of the meaning. The photographs do not provide total visual access. Instead, they reveal the limits of visibility under conditions of secrecy. They show that some forms of power are present precisely because they are kept out of sight.

His later machine-vision projects move the problem from hidden places to hidden systems. In works involving AI training images, facial analysis, and computer vision, Paglen asks viewers to consider a world in which images are made by machines for machines. In this world, the traditional viewer may no longer be central. The image becomes part of an automated chain of classification and control. Paglen’s art therefore does not simply represent surveillance. It interrupts the fantasy that surveillance is invisible, objective, or inevitable.

Hito Steyerl’s work offers another important way to understand the politics of contemporary images. In her influential essay “In Defense of the Poor Image,” Steyerl describes the poor image as a low-resolution, compressed, copied, and circulated image that moves rapidly through digital networks.⁵ Although this essay is not only about surveillance, it helps explain how images lose material stability as they move through systems of circulation. Steyerl’s later work also addresses the relationship between screens, data, war, corporate power, and artificial intelligence. Her practice shows that digital images are never only visual; they are embedded in economic and political infrastructures.

The politics of machine vision also extends into policing and borders. Facial recognition, license-plate readers, body cameras, predictive policing software, biometric databases, and border surveillance systems all use photographic or visual data to manage populations. These technologies are often justified through the language of security, efficiency, and prevention. But they raise serious questions about consent, accuracy, racial bias, privacy, and democratic oversight. Who is watched most intensely? Who is misidentified? Who has access to the data? Who can challenge a system’s decision?

Borders are especially important because they reveal the relationship between visibility and control. Migrants, refugees, and asylum seekers are frequently subjected to biometric identification, surveillance imaging, drone monitoring, and database tracking. The border is no longer only a physical line. It is also a visual and data system. A person may be identified, categorized, and restricted through images long before they encounter a human official. The photograph becomes part of the machinery of mobility and exclusion.

At the same time, surveillance is not experienced equally. Some bodies move through camera systems with relative invisibility, while others are more likely to be stopped, flagged, watched, or misread. Browne’s work on surveillance and Blackness is essential because it shows that contemporary technologies are connected to historical practices of racialized monitoring. Surveillance is not new simply because it uses new tools. It continues older structures of control through updated forms.⁶

This unevenness complicates the idea of privacy. Privacy is often discussed as an individual right, but surveillance is also collective and structural. A person may consent to use a device, app, or platform without fully understanding how their images and data are collected, analyzed, sold, or shared. A community may be surveilled through cameras placed in public spaces without meaningful participation in that decision. A protest may be documented by participants for protection, but the same images may later be used to identify and prosecute them. The camera can witness injustice while also becoming part of a system of repression.

Body cameras provide a useful example of this contradiction. They are often promoted as tools of accountability, especially in response to police violence. In theory, they create a visual record of encounters between police and civilians. Yet body-camera footage is controlled by the institutions that deploy the cameras. What is recorded, when the camera is turned on, who can access the footage, how it is edited, and when it is released are all political questions. The presence of a camera does not automatically produce justice. Images still require interpretation, context, and accountability.

This is one of the central lessons of twenty-first-century photographic literacy: more images do not necessarily mean more truth. Surveillance systems produce vast quantities of visual data, but that data is shaped by institutional interests. A camera may record, but it does not explain. An algorithm may classify, but it does not understand. A database may match, but it may also misidentify. The authority of the image must be questioned precisely because images are now so central to systems of power.

Artists who expose or interrupt these systems help viewers see what would otherwise remain hidden. Hasan Elahi’s long-term project Tracking Transience is one example. After being mistakenly placed under government scrutiny, Elahi began publicly documenting his own movements, meals, airports, hotel rooms, and daily activities. By exaggerating self-surveillance, he turned visibility into a strategy of resistance.⁷ His work asks whether total exposure can disrupt surveillance by overwhelming it with excess information.

Forensic Architecture offers another model. The research agency uses images, videos, satellite data, architectural models, and open-source investigation to examine state violence, human rights abuses, environmental destruction, and war crimes. Their work demonstrates that contemporary images can be used against systems of secrecy when carefully analyzed and contextualized. Photography here becomes part of an evidentiary practice, not as a single decisive image, but as one element in a network of visual, spatial, and temporal data.

This points to a larger tension in contemporary photography. The same technologies that surveil can also expose. A smartphone video may document police violence. Satellite imagery may reveal environmental destruction. Open-source investigators may use social media photographs to reconstruct human rights violations. Drone footage may document state violence or ecological damage. The politics of machine vision is not simply that technology is oppressive. It is that visual systems are contested. They can be used to control, but also to investigate, resist, and testify.

The challenge is to understand who controls the system. An image made by a community member may carry a different political force than an image made by a police surveillance camera. A satellite image used by a corporation to manage extraction is different from a satellite image used by activists to prove illegal deforestation. The technology may be similar, but the power relations around it are not.

In the twenty-first century, then, the question “Who is looking?” has become more complicated. The answer may be a person, a platform, a police department, a drone operator, an AI system, a border agency, a corporation, an archive, or a machine-learning model trained on millions of images. Sometimes no single viewer is looking at all. The image is simply processed.

Photography’s historical relationship to evidence, memory, and representation is therefore being transformed. The photograph still shows, but it also measures. It still records, but it also predicts. It still represents, but it also classifies. It still circulates, but it also extracts data. The camera is no longer only a tool for making images of the world. It is part of the infrastructure through which the world is managed.

This does not mean that photography has lost its human meaning. Rather, it means that human meaning now exists alongside computational uses that are often hidden from view. To study photography in the twenty-first century, students must learn not only how images look, but how they function. Who made the image? Who sees it? Who stores it? Who analyzes it? What data does it produce? What decisions does it enable? Who is protected by it, and who is placed at risk?

The photograph has always been connected to power, but machine vision changes the scale and speed of that connection. A single image may now enter systems of identification, surveillance, classification, and prediction almost instantly. The face, the body, the street, the border, the landscape, and the crowd can all become data. The camera no longer asks only, “What does this look like?” It increasingly asks, “What can be done with this?”

That question defines the contemporary politics of photographic vision. In the age of surveillance, data, and machine seeing, photography is not only about representation. It is about operation. It is about the systems that look through images, act on images, and make decisions from images. To ask “Who is looking now?” is to ask who has the power to turn vision into control.

Footnotes:

1. Harun Farocki, “Phantom Images.” Public, no. 29, 2004, pp. 12–22.
2. Simone Browne, Dark Matters: On the Surveillance of Blackness. Duke University Press, 2015.
3. Kate Crawford and Trevor Paglen, “Excavating AI: The Politics of Images in Machine Learning Training Sets.” AI Now Institute, 19 Sept. 2019.
4. “Machine Visions.” Trevor Paglen Studio, paglen.studio/category/machine-visions/.
5. Hito Steyerl, “In Defense of the Poor Image.” e-flux Journal, no. 10, Nov. 2009.
6. Browne, Dark Matters.
7. Hasan Elahi, Tracking Transience, trackingtransience.net.

---