Making Up Data: Generative AI and the Epistemological and Moral Challenges of Synthetic Image Data

Anna Skarpelis, New York University

The rise of generative AI has sparked troubling developments. Consider when Meta in January 2025 debuted a “sassy black queer” AI assistant called Liv, whose exposure as “digital blackface” saw the company quickly retire the character (Attiah 2025). While the effort aimed to address racial bias, it unintentionally invoked the fraught history of racial transformation and deceit—whether through acts of passing, minstrelsy, or mockery. Most synthetic data is not problematic because it is artificial, but because their mode of creation echoes historically familiar forms of power and exploitation. This article takes a sociological lens to questions of knowledge and power around synthetic data. As a sociologist of science and technology, I study missteps with progressive intention as the ones mentioned above to try and articulate the special challenges that images pose for computational analysis and meaning-making. In the course of this, I pull together the literatures on moral entrepreneurship in AI ethics (Vale 2024), that on agency in machine learning (Stark 2024), the STS literature on translation in the construction of scientific facts (Callon 1984; Latour, Woolgar and Salk 1986; Lynch 1988) and the economic sociology literature on performativity, especially where it concerns person categorization and classification (Espeland et al. 2007; Kiviat 2023; Krippner and Hirschman 2022). This paper has two aims: first, to analyze synthetic data, especially image data, through an STS framework that categorizes it as part of a broader system of generated scientific facts; and second, to examine how power and ethics manifest in generative AI, using examples of progressive initiatives that unintentionally failed.

See paper

 Presented in Session 107. AI Impact on Data Infrastructure