Why Synthetic Data Can Never Be Ethical: A Lesson from Media Ethics

This Dialogue paper argues that the use of synthetic data can never be “ethical.” My argument imports a normative stance from media ethics that “being-ethical-means-being-accountable” (Glasser and Ettema 2008). Building from discourse ethics, this stance positions such ethics as having “the facility to argue articulately and deliberate thoughtfully about moral dilemmas, which in the end means being able to justify, publicly and compellingly, their resolution” (Glasser and Ettema 2008: 512). Crucially, this approach is dialogical and social, necessitating a space open to all affected by relational practices and processes. While the use of synthetic data in commercial institutional contexts may offer workarounds to privacy concerns regarding personally identifiable information (PII) or unpaid user labor––or seem relatively innocuous, as in the case of training computer vision algorithms in video games––this facilitates, as others in this Dialogue section argue in their respective papers, a “fix” or “solutionist” framing that elides ethics and de-politicizes synthetic data. Synthetic data therefore intensifies a pre-existing lack of accountability inherent within automated systems more generally, and through this, entrenches and compounds surveillant practices. In some arenas, the stakes are quite literally life or death, such as in the development of medical AI, and more perniciously, the migration of models from commercial to state deployment in law enforcement and military contexts. Given the foreclosure of thoughtful, articulate, and reflexive inclusive deliberation on the significant moral implications of AI’s vast and ever-growing assemblages, and synthetic data’s role in further mystifying and legitimating its seemingly unbridled development and deployment, I argue that synthetic data can never meet the standard of “ethical” practice.