Paper short abstract:

Discourses of AI promise technologies of visual "understanding" trained on "realistic" images. However there is a complicated relationship between images and truth. We argue that the field has a positivist theory of images divorced from anthropology's understanding of technology and reality.

Paper long abstract:

Discourses of Computer Vision (CV) technologies promise to "transform" how humans and computers interact with their environments through the analysis of "raw perceptual information". However these emerging technologies often rely on a small number of image datasets developed, predominantly, by a handful of elite Western institutions. These image datasets encode a range of social and cultural biases, including western-centric sources, English-language metadata, and underrepresentation of minoritized groups. In response, in recent years the CV discipline has spent significant effort on building image datasets which are larger, more realistic, and more diverse. By doing so, the narrative goes, image datasets can be sufficient proxies of the world that technologies trained on them achieve "visual understanding". However the meaning of each image is assumed to be encoded within the image itself. In adopting these narratives, discourses and practices of CV have constructed theories of images and image interpretation which contrast with understandings of these as cultural artefacts and culturally situated processes. Drawing inspiration from Diana Forsythe's study of the construction of knowledge in Artificial Intelligence, this paper analyses CV discourse to shed light on AI's construction of images, and their relationship with truth. We analyse CV's accounts of image production and normative image ideals, as well as accounts of the relationship between images, truth and reality. We reveal an underlying positivist theory of images and their truth which is divorced from accounts of interpretation and meaning central to anthropology's understanding of technology and the social world.