The Proper Likeness and the Models that Matter

AI systems can now encode and render synthetic likenesses, decoupling visual identity from its bearer. This paper argues that without a coherent philosophical account of what constitutes a "proper" likeness, policy responses to deepfakes and generative imagery will conflate distinct problems. Provides the theoretical groundwork for the harms taxonomy developed in the companion paper with Bariach and McBride.

Develops the concept of "proper likeness" by analogy with Kripke's theory of proper names, arguing that a likeness functions as a rigid designator that fixes reference to a historically understood being rather than a bundle of visual features. Uses encoding and decoding from information theory and semiotics to formalise how likenesses are produced and interpreted, establishing an analytical foundation that extends beyond facial recognition to synthetic imagery generally.