Towards a Harms Taxonomy of AI Likeness Generation

Generative AI systems trained on sufficient images of a person can replicate their likeness without consent, creating a new category of platform-mediated harm. Presents a seven-category taxonomy of harms from synthetic likeness generation and distinguishes generation from distribution as separate vectors requiring distinct policy interventions.

Introduces "indexical sufficiency" to describe the threshold at which a generated image becomes recognisable as a specific person, formalising the relationship between training data and likeness fidelity. Traces how "likeness" has functioned in legal discourse and argues that AI-generated representations challenge existing frameworks by decoupling likeness from any originating act of depiction.