On inscription and bias: data, actor network theory, and the social problems of text-to-image AI models

Abstract

Text-to-image generation platforms are a type of generative artificial intelligence that can produce novel and realistic images from a text prompt. However, these systems also raise social and ethical issues related to the data they rely on. Therefore, this review essay explores how data influence these issues and how to address them using the concept of inscription by Bruno Latour. Inscription is the process of encoding the values and interests of the actors involved in the creation and use of a technology into the technology itself. Using inscription as a theoretical and analytical tool, this work analyzes the data sources, data processing, data representation, and data interpretation of these systems, and reveals how they shape the images they generate and the potential biases and harms they may cause. Thus, this essay offers a new perspective on the ethical discussion of the generative AI models, especially text-to-image models, by bridging the gap between the technical and sociological perspectives on these issues, which has been largely overlooked in the existing literature, and it also provides some novel and practical recommendations for the developers, users, and regulators of these technologies, based on the findings and implications of the analysis.