Unseen things: Computer vision and the uneven geographies of AI failure

AI computer vision systems are increasingly used to extract spatial information from images, yet their performance remains uneven across object types and geographic contexts. This paper argues that such unevenness is bound up in the production of space itself. Through a critical visual methods approach, we analyze the failures of a leading commercial cloud-based object detection system in identifying houses representative of diverse countries and income levels. This empirical case provides a lens for examining how machine perception produces differential geographies of visibility and recognition. We identify two interrelated mechanisms underlying geographic disparities in object detection: the uneven representation of the world within training datasets, and a misalignment between the perceptual logics of deep learning models and the complexity of real-world spatial scenes. Together, these dynamics generate two geographies of AI failure : an analog geography of ‘real space’ rooted in dataset bias, and a digital geography of ‘image space’ shaped by perceptual bias, where models struggle to see spatially continuous, cluttered, or occluded objects. As object detection becomes a leading spatial technology, we argue for a pluralistic machine ontology of perception capable of apprehending whole scenes rather than merely isolating discrete objects. By foregrounding the unseen things of computer vision, the paper contributes to emerging scholarship on the geographies of AI by demonstrating that failure is a constitutive mechanism through which AI's geographies are both produced and made visible.