Why Perceptual Ambiguity Cannot Be Tokenized in Generative AI

Source of Publication

Proceedings of the 31st International Symposium on Electronic/Emerging Art (ISEA2026)

Conference Year

2026

Abstract

Perceptual ambiguity has been a defining resource in modern and contemporary visual art, enabling images to resist closure and sustain perceptual engagement beyond representation. Rather than signaling uncertainty of meaning, ambiguity operates at the level of appearance, where visual relations remain open, reversible, and responsive to attention over time. This paper argues that contemporary token-based generative AI systems cannot enact perceptual ambiguity, even though they may produce images that resemble it. Text-to-image and image-to-image models are built on representational pipelines in which meaning is stabilized in advance of visual formation. Language and images are encoded into semantically structured latent spaces, and synthesis proceeds by resolving uncertainty toward coherent, recognizable outcomes. Within this framework, visual instability or stylistic variation may occur, but it does not persist as a sustained perceptual condition. Through a combined phenomenological and structural analysis of diffusion-based generative systems, the paper examines the asymmetry between encoding and decoding and shows why perceptual reversibility cannot be recovered once analytic stabilization has taken place. Ambiguity does not reside in tokens, nor can it be introduced through descriptive refinement. Its absence in generative imagery marks a conceptual boundary between representational computation and perceptual organization, clarifying both the power and the limits of generative AI for contemporary artistic practice.

Document Type

Conference Proceeding

Publication Date

11-6-2026

Author First name, Last name, Institution

This document is currently not available here.

Share

COinS