Towards a Harms Taxonomy of AI Likeness Generation
Ben Bariach, Bernie Hogan, Keegan McBride
TL;DR
The paper investigates the emergence of likeness generation in generative AI, defining likeness through a legal lens and assessing harms via a seven-category taxonomy grounded in meta-analytic online-harms literature. It analyzes the technical underpinnings, highlighting diffusion-based foundation models and personalization techniques such as DreamBooth and LoRA, and argues that likeness generation relies on latent identity representations rather than mere image Warping, enabling broader, harder-to-trace use cases. Four mitigation-guiding criteria—indexical sufficiency, generation vs distribution, denoted vs contextual harms, and benign vs harmful prompts—are proposed to steer governance across the likeness supply chain. The work underscores the need for ongoing research and policy development to balance legitimate creative uses with protections against reputational, sexual, CSAI, security, representational, misinformation, and illocutionary harms in a multimodal AI future.
Abstract
Generative artificial intelligence models, when trained on a sufficient number of a person's images, can replicate their identifying features in a photorealistic manner. We refer to this process as 'likeness generation'. Likeness-featuring synthetic outputs often present a person's likeness without their control or consent, and may lead to harmful consequences. This paper explores philosophical and policy issues surrounding generated likeness. It begins by offering a conceptual framework for understanding likeness generation by examining the novel capabilities introduced by generative systems. The paper then establishes a definition of likeness by tracing its historical development in legal literature. Building on this foundation, we present a taxonomy of harms associated with generated likeness, derived from a comprehensive meta-analysis of relevant literature. This taxonomy categorises harms into seven distinct groups, unified by shared characteristics. Utilising this taxonomy, we raise various considerations that need to be addressed for the deployment of appropriate mitigations. Given the multitude of stakeholders involved in both the creation and distribution of likeness, we introduce concepts such as indexical sufficiency, a distinction between generation and distribution, and harms as having a context-specific nature. This work aims to serve industry, policymakers, and future academic researchers in their efforts to address the societal challenges posed by likeness generation.