Invisibility of Computer Vision: Crafting Machine Vision through Human Eyes

Invisibility, where image data escapes data scientists’ observation, is a critical yet overlooked dimension shaped by human perception in AI alignment and reliability. Drawing on research from nine months of ethnographic fieldwork in a computer vision (CV) laboratory, this paper gives thick descriptions on how scientists train models with their “social body.” We explore the technical characteristics of invisibility and its influence on the material infrastructure and experimental processes involved in AI knowledge production, including dataset classification, GPU allocation, debugging, and validation. Further, we show how data scientists address invisibility through tacit knowledge, inscriptions, and collaboration with science community and social survey. By emphasizing the role of human visual perception in “crafting” machine vision, this paper argues that machine learning is inherently human-centered, involving the collective consultation of different social groups in a sensory way.