Comics Pages as Audit Devices for Generative AI: A Gutter/Ghost Framework

Comics are sites for experimenting with generative AI in human–AI co-creation. Even as debates around these systems intensify, they are often treated as neutral tools rather than as socio-technical infrastructures to be interrogated. Building on my work on educational and data comics as epistemic interfaces and page-level infrastructures, I propose comics pages as testbeds for text-to-image (T2I) automation. The paper argues that generative models can produce plausible panels but struggle with page-level coherence: tracking character identity, stabilising environments, and sustaining spatial, causal and multilinear relations across panels and pages. I read these failures as structural, rooted in sequential, autoregressive architectures—which generate each element step by step from what comes immediately before—and in training regimes optimised for local plausibility and linear associations rather than for non-linear, multi-image relational reasoning. Because comics pages pack verbal–visual units on a surface, while being ingested by models as flattened images, they are well suited as test environments that make these limits legible. I frame this through the Gutter/Ghost framework. “Gutter” (the inferential space between panels) names points at which local, page-level configurations break down. “Ghost” denotes hidden labour—prompting, selecting, assembling, re-captioning, redrawing in loops that repair and reconfigure AI outputs into a cohesive whole. Methodologically, I outline a framework combining page-level analysis of comics layouts as technical configurations with prototypes of hybrid human–AI workflows. I argue that AI-assisted comics can function as audit devices that render visible the infrastructures, limitations and ghost work underpinning regimes of visual automation.