The project examines the changing epistemic status of images under generative AI. While critical work has largely focused on consumer-facing text-to-image and video generation, the thesis turns to technical, scientific and industrial image operations, where images are treated as incomplete evidence from which hidden structures, missing information, world knowledge, intelligent action, or possible futures can be inferred. The project argues that such techniques do not simply extract information already contained in images, but negotiate local observations with learned statistical priors, making the boundary between reconstruction, prediction and synthesis increasingly unstable. Situated between media theory, Bildwissenschaft, history and philosophy of science, and critical AI studies, the thesis asks which conceptions of images make these operations technically possible and epistemically plausible. In turn, it reads architectural debates in the fields of AI and machine learning as latent theories of media: presupposing, challenging or reinforcing mediatheoretical assumptions about digital images.