Trevor Paglen’s New Book Argues Images Now Work for Machines First
What if the most important thing an image does today is not to be seen, but to trigger a system? That is the premise behind How to See Like a Machine: Images After AI, the new book by American artist Trevor Paglen (b. 1974), which extends a long-running body of work on surveillance, computer vision, and the hidden infrastructures of contemporary life.
Paglen’s argument is blunt: the familiar human-centered idea of the image as representation is no longer enough. In his account, images increasingly operate as “activations” inside technical and cultural circuits, where they prompt automated, preconscious, or affective responses. The shift, he writes, is from asking what an image says to asking what it does.
He frames this change around two upheavals in the history of seeing: computer vision and generative AI. Computer vision, he argues, collapses the visual field into vectors and mathematical abstractions. Generative AI then uses those abstractions to shape reality itself. The result is what Paglen calls “machine realism,” a world in which machine-readable images are built to influence behavior, extract value, and normalize surveillance.
The book grounds that claim in ordinary examples. Grocery self-checkout cameras flag suspected shoplifting. Samsara’s truck-monitoring systems watch drivers for safety violations. ImageNet, described as a default training set for numerous AI models, underwrites reductive taxonomies that feed facial recognition. Microscan automates packaging, shipping, and logistics for industries including electronics and pharmaceuticals. Even social media, Paglen notes, measures dwell time, shares, comments, and biometric responses to refine algorithmic targeting.
Paglen also places the present moment in a longer intellectual and political lineage. He acknowledges predecessors such as Vilém Flusser, Paul Virilio, Harun Farocki, and Hito Steyerl, while arguing that today’s machine-to-machine seeing has intensified the stakes. Images now look back at users, adapt to them, and help systems learn from them.
That concern runs through Paglen’s earlier project ImageNet Roulette, which exposed the biases embedded in facial recognition by mislabeling uploaded images, including those of Paglen and Crawford. In the new book, he extends that critique into the language of cognitive warfare, linking contemporary information operations to older programs such as MKUltra.
The larger implication is unsettling but clear: images are no longer only cultural objects. They are operational tools, embedded in systems that shape perception, labor, and power at scale.
